perf(kernel): make generic pad replay incremental

[shinaoka]

Summary

• precompile checked valid pad-copy intervals and incremental fill/copy cursors
• remove per-element rank scans and repeated bounds checks from generic initialized/uninitialized replay
• add frozen rank/layout benchmarks and correctness, threshold, and allocation coverage

Closes #240
Part of #213

Performance

Pinned EPYC CPUs 1-4, four Rayon workers, setup outside timing:

• medium serial compact rank 2/4/8: 2.56x / 4.84x / 8.43x
• medium 4T compact rank 2/4/8: 1.35x / 2.30x / 3.50x
• medium negative crop / nonunit serial: 2.59x / 2.10x
• large generic cells: 1.38-8.44x
• all 40 frozen generic cells improved with non-overlapping intervals
• unchanged rank-one medium/large controls remained stable

Full baseline/candidate evidence is in docs/worklogs/2026-08-24-issue-240-pad-fallback.md.

Verification

• cargo fmt --all -- --check
• focused default: 85 passed
• focused parallel: 96 passed
• rank-8 generic initialized/uninitialized replay: zero allocations
• source contracts: 6 passed
• cargo test --workspace: 915 passed, 9 ignored
• cargo test --workspace --features parallel: 989 passed, 9 ignored
• cargo check -p strided-kernel --features parallel
• deterministic repository-rules review: pass
• modified-file coverage: static_indexing_plan.rs 81.93% (threshold 80%)

The global coverage script also reports unchanged pre-existing deficits in reduce_view.rs and hptt/execute.rs; neither file is touched here.

Review gates

• design review: reviewer-flash high/read-only — Correct-to-merge at 303720d
• exact final diff review: reviewer-flash high/read-only — Correct-to-merge at e1f01cdd