Phase 1 — Port BT2's GroupWalk SA-range traversal

[k1sauce]

Part of #1. Blocked by #2 (need root cause confirmed first).

Scope

BT2 walks SA-range hits in a specific deterministic order via `GroupWalk` (`vendor/bowtie2/group_walk.h`). The order matters because BT2's `-D` budget can fire mid-traversal, leaving a specific subset of hits explored. Our current `collect_prioritized` in `crates/rusttie-align/src/align.rs:531` uses a simpler sort by SA-range size — different traversal order, different alignments discovered.

If Phase 0 confirms GroupWalk is (part of) the cause, port it.

Deliverables

• New module `rusttie-index::groupwalk` mirroring the public API of `group_walk.h`.
• Per-element validation: feed a known SA range, assert output sequence matches BT2's byte-for-byte. Use the instrumented BT2 binary from Phase 0 — Spike: identify exact BT2 pool divergence root cause #2.
• Replace `collect_prioritized` to use `GroupWalk` (gated behind `RUSTTIE_USE_GROUPWALK=1` initially for A/B safety).
• Single-end chr22 validation: position / CIGAR / AS agreement stays ≥99.6%.
• Joint-descent uses `GroupWalk` once gated.

Exit criterion

• Single-end metrics unchanged on chr22 (regression gate).
• Pool-diff diagnostic shows our pool composition shifting toward BT2's for the Phase 0 sample reads.

Risks

• BT2's `GroupWalk` uses a deterministic PRNG seeded from read+position. Reproducing the same PRNG sequence requires matching BT2's RNG implementation (`vendor/bowtie2/random_source.h`).
• The walk can have side-effects on a shared cache structure (`AlignmentCacheIface`) that we'd need to model or stub.

Files referenced

• `vendor/bowtie2/group_walk.h`
• `vendor/bowtie2/random_source.h`
• `crates/rusttie-align/src/align.rs:531` (current collect_prioritized to replace)