enable_fast_train() routes QLoRA training through the grouped NF4 kernel via
grouped-nf4-gemm>=0.2.4's differentiable nf4_qlora surface. Verified on
Qwen3-30B-A3B at 48 layers with expert offload + gradient checkpointing:
held-out eval 0.337073 (ref) vs 0.340271 (fused), 0.95% apart against a
registered 0.05 band; peak 6.87 GB vs 9.09 GB reference (0.756x); 192 frozen
expert tensors bit-identical over 16.3 GB hashed with a byte-flip positive
control.

Also: --help no longer loads a model, __version__ no longer drifts from the
distribution, and the train-mode fast path warns instead of silently
no-opping.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>