Refactor pybigtools values() rasterization

[nvictus]

This PR reimplements the binned/base-level rasterization in pybigtools around a cleaner API, with explicit semantics distinguishing the inclusion of uncovered bases (resolves #94) in summary statistic calculations and the post-rasterization filling of empty bins or positions. Internally, it also introduces a clean sweepline-based iterator for generating depth-of-coverage intervals from BigBed entries (the prerequisite for BigBed rasterization) and exposes the full set of internal summary stats through BBIReader.values().

1. Uncovered-base inclusion/exclusion in statistics and postfill

Previously, the missing argument served the purpose of post-rasterization fill-in but not inclusion of uncovered bases in summary calculations, and it's naming was ambiguous.

There is demand for summary statistics that consider uncovered bases as zeros (summarize over the entire bin) rather than only over covered bases, which is not yet exposed via values() #94. UCSC exposes some of these with names like mean0 in averageOverBed but this is all somewhat obscure and many people don't even realize that plain mean excludes uncovered bases.

The goal here is to make both functionality possible via two arguments:

• uncovered: float | None -- determines whether uncovered bases are included in statistics. By default (None) they are excluded and any empty positions or summary bins are returned as NaN. You can opt-in to treating all uncovered bases as having value 0.0 via uncovered=0.
• fillna: float | None -- replaces missing and determines how any in-bounds NaN values are filled after rasterization. The previous default behavior was to replace them with 0. The new default fillna=None is to leave them as is (which is coincidentally what pyBigWig does). Users can opt in to the old behavior with fillna=0, which is much more ergonomic than post-filling using numpy primitives.

Through the combination of summary and uncovered, we expose the equivalents of mean0, std0, min0, max0. As a bonus, you can insert any finite background value for uncovered bases you like.

This PR also exposes the full set of internal summary stats, including sum, sum_squares, bases_covered (number of bases covered by features), and bin_covered (fraction of the bin covered by features). The latter two are not affected by the uncovered argument by definition.

Deprecation cycle

Default behavior change: when fillna is not passed, empty bins / uncovered positions are now returned as NaN instead of being filled with 0. A DeprecationWarning is raised on the first call without an explicit fillna to flag the change for existing users; pass fillna=0 to keep prior behavior or fillna=None to silence the warning.

In order to emit warnings, a thin Python wrapper was introduced around the Rust BBIReader to detect "user did not pass fillna" (a distinction PyO3 can't make at the Rust boundary) and emit the DeprecationWarning. It can be removed in a future release when the deprecation is done.

API changes (Python BBIReader.values())

Parameter	Status	Semantics	Default
uncovered	new	Per-base value for uncovered bases used during the summary computation. None (default) excludes them (classical stats); 0.0 reproduces UCSC's *0 stats; any finite value is a well-defined generalization.	None (exclude)
fillna	new	Post-rasterization fill for in-bounds NaN positions/bins. OOB positions are not touched by fillna.	None (keep NaNs)
missing	deprecated	Accepted as an alias for fillna with a DeprecationWarning.	old default was 0.0; now opt-in with fillna=0
oob	unchanged	Fill for positions outside the chromosome's valid range.	NaN

2. UCSC parity in zoom summary interpolation (ceil + rescale normalization)

The zoom-interpolation code path now matches UCSC's bbiSummarySlice normalization: per-bin bases_covered is rounded up to the next integer, with sum and sum_squares rescaled by ceil(bc)/bc so that mean = sum / bases_covered is invariant. The ceil direction preserves the invariant that bases_covered ≥ 1 whenever any data overlaps the bin (a floor would erase data near zoom-record boundaries). std is not algebraically preserved under the rescale, but matching UCSC's reported std requires applying the same step.

3. Coverage iterator

CoverageIterator, a 1-D sweep-line over endpoint events that converts sorted BED entries to a run-length-encoded coverage track (matching bedtools genomecov -bg semantics).

Tests

• 31 Rust unit tests (up from 4): CoverageIterator (15 cases including sort-violation, malformed-input, stress, batching), fill_binned (8 cases covering every Summary variant plus the integer-bin-width edge case).
• 715 Python tests (up from ~580): parity vs pybbi for sum/mean/ std/min/max/bin_covered/bases_covered across many start/end/bin combinations and both base-level and binned modes. Adds zoom-interpolation parity tests (test_bbi_consistent_binned_zoom_bw/bb) againstpybbi. Plus tests for fillna/uncovered semantics, the deprecation warnings, the arr= parameter, invalid arguments, and fully out-of-bounds queries.
• pyproject.toml: filterwarnings ignores DeprecationWarning from pybigtools during the test run; deprecation tests opt back in locally via pytest.warns(...).

Migration notes for users

# old
bw.values("chr1", 0, 100, bins=10, missing=-1)                # → DeprecationWarning
bw.values("chr1", 0, 100, bins=10)                            # → DeprecationWarning, NaN on empty (changed)

# new
bw.values("chr1", 0, 100, bins=10, summary="mean", uncovered=0)  # translates to "mean0"
bw.values("chr1", 0, 100, bins=10, fillna=-1)                    # = old missing=-1
bw.values("chr1", 0, 100, bins=10, fillna=None)                  # opt in, no warning
bw.values("chr1", 0, 100, bins=10, fillna=0)                     # opt back into old default

[jackh726]

I'm trying to remember - I had thought a bunch about this, and there was some reason (or some difference) in choosing missing versus the fillna parameter. I'll have to think a bit about it and try to recall. Maybe splitting into uncovered and fillna fixes this.

[nvictus]

missing is the name I used in pybbi which I now regret given the confusing semantics. Whatever we do here I may backport there.

[nvictus]

Fixed the nightly build issue @jackh726.