Split numerical kernels into ordered stages

[7etsuo]

Summary

• Split layernorm, matmul backward, attention backward, AdamW, and categorical sampling into small named stages.
• Preserve numerical expression order, loop order, accumulation semantics, OpenMP ownership, and RNG consumption.
• Add regressions for nonzero gradient destinations, causal scratch suffixes, rejected optimizer state, two-row normalization, and one-id sampling.
• Synchronize the source walkthroughs, lab counts, source-line count, and evidence record.

Compatibility

A 50-step GCC 13.3.0 serial training witness against merged commit a4dfaae produced a byte-identical checkpoint. Sampling produced byte-identical stdout and stderr. The recorded checkpoint SHA-256 is 4adcaab12765e7d2b0802851a2bbb64c723c90263dea725c2f5cbe7d0f1627e7.

Validation

• make OPENMP=0 check-all
• make OPENMP=1 check-parallel
• make check-sanitizers
• Clang focused backward, optimizer, and sampling checks
• make check-book
• python3 scripts/check-graph.py
• evidence manifest and compatibility byte comparisons
• two independent merge-readiness reviews

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.