[zhangzhenghao]

Summary

Depends on #1398 (FP16 tensor storage). This stacked PR targets pr-1398-fp16-storage, an exact mirror of #1398's storage-only head, so its diff contains computation changes only. Merge #1398 first, then retarget this PR to master.

• Dispatch FP16 operators through transient FP32 conversion and existing FP32 kernels; round outputs to the first dispatched input's dtype, preserving existing Add/Mul operand reordering.
• Preserve original graph connectivity and retain FP32 kernel outputs strongly for backward operations; keep gradients FP32 and support graph-free inference.
• Update FP16 parameters with SGD/Adam using transient FP32 values, rounding into existing FP16 backing storage to preserve aliases. Keep Adam moments FP32; no persistent master weights.
• Support FP16 NormalInit and cover primitive forwards/backwards, matrix transpose variants, mixed operands, chained/shared graphs with forced GC, inference, optimizer multistep updates/weight decay/aliasing, and initialization.

Verification

Passed locally with system Go 1.27.3, without modifying go.mod/go.sum:

• go test ./common/nn ./protocol -count=1
• go test -race ./common/nn ./protocol -count=1
• go test -tags=noasm ./common/nn ./protocol -count=1
• go vet -composites=false ./common/nn ./protocol
• git diff --cached --check
• Independent review: no blocking security or logic issues; split integrity verified.

The resulting source tree exactly matches the former combined #1398 head 5e39e1a4a3cff412df6f6b12683f42ea96b30692 (tree adea68644abaec590de64501a76736ae3ac40b13). No computation behavior was changed by the split. GitHub Actions currently filters pull requests to master / release-**, so this stacked PR does not trigger remote CI while targeting pr-1398-fp16-storage. Local verification above passed; remote CI must run after retargeting to master once #1398 merges.