Sink a filter that holds a projection into NdFilterExec (#397)

[robinskil]

What changes does this PR make?

NdFilterPushdown now sinks a FilterExec that holds a projection. The conjuncts go
into NdFilterExec. The projection becomes an NdProjectionExec below the broadcast.

Why do you make these changes?

Closes #397. DataFusion adds a projection when the select list is narrower than the
predicate. count(*) gets an empty projection. The rule refused such a filter, so the
plan broadcast every column onto the full grid first.

How do you make these changes?

A residual conjunct reads the columns that the projection drops. So the projection stays
with the residual filter above the broadcast. The rule also keeps a filter that holds a
fetch, because the nd filter has no cap.

Release run on gridded-example.nc, 2.32 M cells, 714.1 K rows survive:

query	broadcast before	broadcast after
SELECT lat … ST_Within(…)	8.30 ms, 2.32 M rows	2.08 ms, 714.1 K rows
count(*) … ST_Within(…)	5.33 ms, 2.32 M rows	10.08 µs, 0 columns

Does this PR change a public interface or a configuration?

No.

How do you test these changes?

Four unit tests in nd/mod.rs: a narrow projection, an empty projection, a residual
conjunct, and the fetch guard. Each test compares the rewritten plan against the rows
of the original plan.

cargo clippy -p beacon-datafusion-ext --lib --tests
cargo test -p beacon-datafusion-ext --lib

I also checked the queries against a runtime with the nd pipeline off. The rows, the
count(*) and the sum() agree.

Checklist

• The title of the PR describes the change, and it ends with the issue number if one exists.
• The code follows the style of the project, and cargo fmt --all --check passes.
• cargo clippy --workspace --lib --bins --tests reports no new warnings.
• cargo test --workspace --lib --bins --tests passes.
• The new code has tests, or the PR explains why it does not.
• The documentation, the README, and the CHANGELOG show the change.
• The PR contains no unrelated changes.

Notes on the two open boxes:

• cargo fmt --all --check does not pass on main today. The new code follows the
  style of the file.
• The CHANGELOG bullet on N-dimensional execution already states this behaviour, and
  the feature is unreleased.
• PR Apply the whole pushed-down projection in the ODV and BBF scans (#382) #398 adds beacon-core/tests/nd_pipeline_filters.rs. Its test
  a_narrowing_select_list_keeps_the_filter_above_the_broadcast records the old
  behaviour, so it needs an update there.

Was this patch authored or co-authored using generative AI tooling?

Assisted-by: Claude:claude-opus-5

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 82.46%. Comparing base (2f8f748) to head (6ff9a07).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #401      +/-   ##
==========================================
+ Coverage   82.42%   82.46%   +0.04%     
==========================================
  Files         363      363              
  Lines       59366    59520     +154     
==========================================
+ Hits        48931    49086     +155     
+ Misses      10435    10434       -1     

Files with missing lines	Coverage Δ
beacon-db/beacon-datafusion-ext/src/nd/mod.rs	98.15% <100.00%> (+0.44%)	⬆️
...eacon-db/beacon-datafusion-ext/src/nd/optimizer.rs	89.09% <100.00%> (+1.97%)	⬆️

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.