perf: use native words in IndexedSet and hoist forbid lookups

[baszalmstra]

IndexedSet is backed by a growable Vec<usize> bitset, using the native machine word size while preserving the existing set API.

The encoder also resolved a solvable's package name and looked up its pending forbid-clause bucket once per candidate. Candidates in one version-set candidate list share a package name, so this now happens once per list. The allow_multiple check and forbid bookkeeping move out of the inner loop as well.

What changed

• Store IndexedSet bits in native machine words and remove the unused bitvec dependency.
• Resolve forbid-target names once per candidate list instead of once per candidate.
• Assert in debug builds that every candidate list has one package name.
• Add regression coverage for word boundaries, sparse and large indices, duplicate insertion, unions, empty candidate lists, locks, exclusions, and allow_multiple.

Benchmarks

Benchmarked with a 1,134,320-record conda-forge snapshot using 351 fixed-seed, precomputed workloads. Both builds used the same requirement sequence, a 60-second per-solve timeout, and produced identical outcome categories: 129 solved and 222 unsatisfiable.

metric	main	this PR
total solve time	130.25 s	106.07 s
mean	371.1 ms	302.2 ms
median	257.7 ms	202.1 ms
p95	909.4 ms	729.4 ms
max	3.78 s	3.43 s

This reduces total solve time by 24.18 s (18.6%), or 1.23x throughput.

339 of 351 problems are faster. The largest improvement is 358 ms; the largest regression is 25 ms.

Testing

• pixi run -e test-rust cargo fmt --check
• pixi run -e test-rust cargo test --all-features

[dralley]

@baszalmstra What is the impact on memory usage?

[baszalmstra]

I didnt measure! But I expect very little since this is only used with dense layouts.