Add INT_W8A8_CONVROT: W8A8 quantization with Hadamard rotation

[dxqb]

Summary

# Prototype knobs - env-driven, no UI/config plumbing (see PLAN in-an-new-branch-reflective-cocke).
# Block size for the group-wise Hadamard rotation. Not empirically validated; 128 is a common
# group size in block quantization literature, picked as a starting point to sweep from.
CONVROT_BLOCK_SIZE = int(os.environ.get("CONVROT_BLOCK_SIZE", "128"))
# Isolates the untested part of this experiment: whether quantizing the backward grad-output
# to int8 (matching plain int8w8's backward) hurts LoRA gradient quality vs keeping it in
# bf16. Default matches int8w8 so the two paths are directly comparable.
CONVROT_BF16_DY = os.environ.get("CONVROT_BF16_DY", "0") == "1"

Test plan

• pre-commit run --all-files passes
• Launched the affected UI or script and exercised the change
• Tested with at least one real preset / config when relevant (note which: ____)

AI assistance

• Early AI prototype — opened for discussion, not ready for review