Optimize WeightedIndex weight lookup and iteration

[pnn64]

WeightedIndex::weights() can avoid repeated vector growth by reporting its remaining length. Weight lookup can also use a simpler checked access path.

Solution

Implement size_hint and ExactSizeIterator for the weight iterator, allowing collection to reserve enough capacity up front.

Replace the three-way index comparison in weight() with checked slice access and a fallback for the final weight. Returned values, clone/subtraction order, and out-of-bounds behavior are preserved. The sampling algorithm and RNG calls are unchanged.

Tests, benchmarks, and the changelog are updated.

Benchmarks

On a Xeon E5-2696 v4, using rustc 1.98.0 on Windows in release mode.

Each operation below processes 1,024 weights; distribution construction is excluded.

Measurement	Before	After	Change
Collect u32 weights	2,969.77 ns	1,218.95 ns	-59.0%
Collect f64 weights	3,175.76 ns	1,507.99 ns	-52.5%
Sum u32 weights	1,679.96 ns	1,009.58 ns	-39.9%
Sum f64 weights	1,858.66 ns	1,110.69 ns	-40.2%
Extend reused buffer, u32	1,848.52 ns	1,129.23 ns	-38.9%
Extend reused buffer, f64	1,840.83 ns	1,121.59 ns	-39.1%
Growth reallocations per collection	8	0	Eliminated
Initial allocations per collection	1	1	Unchanged

Timing: Criterion, 100 samples, 0.5-second warm-up, and 2-second measurement target. Each version ran twice in before/after/after/before order; times are averages of the two estimates. Allocation counts were measured separately.

Across 36 cases, 27 showed improvements beyond the measured timing uncertainty. The other nine had overlapping before/after confidence-interval envelopes; none showed a resolved regression.

Testing and linting

• Default library tests: 100 passed.
• Allocation-only library tests: 84 passed.
• Iterator coverage includes exact remaining lengths, cloning, partial consumption, and repeated exhaustion.
• Lookup coverage includes out-of-bounds indices and usize::MAX.
• Clippy with warnings denied and formatting checks passed.

🤖 AI disclosure: I used Codex with GPT-6 Astra at xhigh reasoning effort to help identify the optimization and construct the benchmark. I have reviewed and understand the resulting implementation.

[dhardy]

Thanks for the optimisations!

Code looks good to me with some minor suggestions below.

The CHANGELOG needs to link to this PR.