strided-kernel: incremental offsets in erased reduce and indexed replay hot loops

[shinaoka]

Part of #199 (Phase 3, performance). All file:line references are pinned to origin/main at 9da9b9f.

Per-element offset recomputation in tensor-sized loops, against the layout rule (avoid per-element flat-to-multi-index decoding; hot loops must not repeat post-validation checks):

• erased.rs:3378 and :3436 (erased axis reduction, serial and parallel): checked_strided_offset(...)? per element, a rank-length checked mul/add chain, plus src_idx rebuild per output element; runs dest_total * reduce_total times.

• Same pattern across static_indexing_plan.rs (:497, :528, :611, :661, :713, :763) and gather_plan.rs (:373, :453, :688, :787, :976, :1075, :1371).

• erased.rs:2376 raw_any: full div/mod flat-to-multi decode per element.

• Hoist validation, precompute stride tables, and step offsets incrementally (the elem_loops! traversal in kernel.rs:244 and advance_col_major_index already do this correctly; reuse them).

• Benchmark before/after across sizes and ranks per the repository benchmark rules.