perf(monorepo): reuse the tag index ancestor set instead of walking twice

[BryanFRD]

run_release_logic walked HEAD's full ancestry twice, and threw the first walk away.

build_head_ancestors (mod.rs:150) and TagIndex::build run the byte-identical revwalk, and TagIndex already exposes the result as pub ancestors. On top of that, all four reads of inputs.head_ancestors (plan.rs:128/160/183/197) live in the else of if let (Some(idx), OrphanedTagStrategy::Warn) = (inputs.tag_index, strategy) — and Warn is #[default] while TagIndex::build normally succeeds. So in the default configuration the first walk was dead work: nothing read it.

src/query.rs:229 already does Some(&idx.ancestors); this path simply missed the same reuse.

Now the index is built first and its ancestor set is passed through; the standalone walk only runs as a fallback when the index is unavailable (which is also the only case where the else branches can fire).

Measured — A/B on the real CI fixture

mono-large (200 packages × 10k commits), generated from benchmarks/fixtures/definitions/mono-large.json with the Fixtures generator and the same ferrflow.json the benchmark harness writes. check --jobs 1, cache cleared before every run, same binary/machine, n=3:

	run 1	run 2	run 3	median
before	2722 ms	2568 ms	2551 ms	2568 ms
after	2069 ms	2078 ms	2137 ms	2078 ms

−490 ms, −19%.

This is not a benchmark-only win. The cache key is head + tags_hash + config_hash (cache.rs:42-56), so every push invalidates it — a one-run-per-push CI job is always cold. Cold numbers are the real numbers.

875 tests pass, clippy -D warnings clean. No observable behaviour change: the ancestor set is identical, only its provenance changes.

Closes #682