Add native DFlash repository LoRA pilot

[MarinaMackay]

Native DFlash Repository LoRA Pilot on Two RTX 4090 GPUs

Summary

This change adds an end-to-end native DFlash training and evaluation path for
repository-specific speculative decoding on a two-RTX-4090 machine.

The implementation includes:

• distributed Stage 1 target LoRA training and evaluation;
• target-adapter merge and frozen-target teacher generation;
• teacher repair, quality filtering, and ModelOpt conversation export;
• native DFlash launch configuration with sequence length 4096, block size 8,
  512 anchors, five drafter layers, D-PACE, and self-logit distillation;
• parameter-efficient DFlash LoRA injection, serialization, and restoration;
• exact online greedy acceptance evaluation;
• deterministic disjoint repository train/evaluation splitting;
• reproducible sweep and final holdout scripts.
• a scoped ModelOpt compatibility patch for the tested PyTorch 2.5 DDP
  environment and trusted local checkpoint resume.

Results

On a deterministic xarray 120/30 train/holdout split with zero example-ID
overlap, repository DFlash LoRA improves mean accepted draft tokens from
0.0218 to 0.1939, an 8.88x increase. The adapter trains 7,454,720
parameters while the repository-adapted target remains frozen.

The full native DFlash configuration completes on two RTX 4090 GPUs without
OOM or NaN at approximately 19.4-19.7 GiB per GPU.

Validation

• Python compilation for all new and modified training modules
• Shell syntax validation for all launch scripts
• JSON parsing for both training configurations
• Dry-run validation for Stage 1 and native DFlash launch construction
• Remote two-GPU generic DFlash training through 1,500 steps
• Adapter save/load evaluation on held-out repository prompts
• Verified zero train/evaluation ID overlap

Detailed measurements and limitations are documented in
NATIVE_DFLASH_2X4090_RESULTS.md.