feat(train): GPU-aware MFU, warmup MFU fix, experiment integrity log

[eli-labz]

Three targeted improvements to train.py:

fix(train): GPU-aware peak FLOPS for accurate MFU on non-H100 GPUs (#547)

• Add GPU_PEAK_FLOPS dict mapping (compute_cap) -> peak BF16 FLOPS
• Covers V100, A100, RTX 3090/A10G, L4/L40S/RTX 4090, H100, B200, RTX 5090
• Falls back to H100 value (989.5e12) for unknown GPUs
• H100_BF16_PEAK_FLOPS remains as the resolved scalar for backward compat

fix(train): warmup off-by-one in steady_state_mfu accumulation (#556)

• Training skips timing for steps 0..10 inclusive (11 steps, not 10)
• Denominator corrected from (step - 10) to (step - 11)

feat(train): experiment integrity log to detect metric gaming (#599)

• Add log_integrity() function using stdlib only (hashlib, datetime)
• Writes one line per run to integrity.log binding val_bpb to:
  • sha256[:16] of train.py and prepare.py source files
  • sha256[:16] of first 1 MB of final model weights
  • real optimizer step count from optimizer internal state (not loop var)
  • wall-clock training seconds
• Detection only, never raises, wrapped in try/except so it can never crash a run; integrity.log should be gitignored like results.tsv