Specialize TupleWindows fold

[CollinYuanjieRen]

Related to #755.

Initialize the first window once, then delegate to the inner iterator's fold. Keep the width-one no-clone path and stop immediately when initial collection fails. Add regressions for partially consumed and non-fused inputs, clone/consumer panics, and callback/iterator/window drop order against the default next loop.

Short Criterion comparison on Rust 1.95.0 / Linux aarch64, 4,096 u64 values, 20 samples with 1 s warm-up and measurement (means):

Input / window width	Specialized	Default loop
Slice / 1	0.454 µs	0.467 µs
Slice / 2	0.472 µs	0.478 µs
Slice / 3	0.550 µs	0.548 µs
Slice / 4	0.756 µs	0.683 µs
Two chained slices / 2	0.491 µs	1.525 µs

The chained case improves about 3.1×; flat slices mostly do not benefit, and width 4 regresses about 11% in this run.

Validation: cargo test --test specializations tuple_windows, cargo fmt --check, and warning-free Clippy for the alloc-only library and changed test/benchmark targets. Benchmark: cargo bench --bench fold_specialization -- tuple_windows --sample-size 20 --measurement-time 1 --warm-up-time 1.

[phimuemue]

TupleWindows is on the way out. Can you deprecate tuple_windows (we have array_windows instead)? If you want, you can try specializing Array windows::fold.