Make generic gather replay incremental

[shinaoka]

Summary

• precompute private generic-gather replay steps/resets for destination, window, and index-batch offsets
• decode once for serial replay and once per parallel worker range, then advance state incrementally
• keep index-vector contributions fresh and data-dependent per element while removing full rank scans and checked-offset reconstruction
• validate signed layout spans and replay deltas, including negative strides
• preserve the compact rank-one fast path and all public APIs
• add generic rank/layout Criterion coverage and ground-truth initialized/uninitialized serial/parallel tests

Fixes #237
Refs #213

Performance

AMD EPYC 7713P, release Criterion threshold profile, topology/load-gated 1T/4T runs. At N = 262,144:

case	baseline	candidate	speedup
rank 4 serial	15.737 ms	2.0825 ms	7.56x
rank 8 serial	27.882 ms	2.4558 ms	11.35x
rank 4 max_threads(4)	1.6719 ms	0.6305 ms	2.65x
rank 8 max_threads(4)	2.6332 ms	0.6965 ms	3.78x
non-unit rank 2 serial	10.418 ms	1.8255 ms	5.71x
negative rank 2 serial	10.378 ms	1.8131 ms	5.72x

All 40 generic cases improved with p < 0.05; all predeclared gates passed. The full 48-case baseline and candidate tables, CIs, affinity/noise gates, and CCD control-shift disclosure are in the worklog.

Work log and review gates

• docs/worklogs/2026-08-24-issue-237-generic-gather.md
• pre-implementation design review: reviewer-flash Correct-to-merge at 9dfdf05
• safety/semantic preflight: no Critical/Important findings at d95a387
• exact-final review: Correct-to-merge at 1be7ef6208b62e1aafb7be09e35346a5262e105b

Verification

• focused default gather/uninitialized tests: 79 passed
• focused parallel gather/uninitialized/policy tests: 88 passed
• cargo check -p strided-kernel default + parallel: pass
• cargo fmt --all -- --check: pass
• cargo test --workspace: 904 passed, 9 ignored
• cargo doc --workspace --no-deps: pass
• deterministic repository-rules review: pass, no findings
• repository-rules review script: 83 passed
• modified gather_plan.rs coverage: 82.95% (repository threshold 80%)

The local all-workspace coverage report also identified three pre-existing unmodified files below their configured thresholds; this PR does not alter them. Hosted CI remains authoritative for the exact coverage gate.