[WIP] feat: add DFlash2 draft serving support

[torchspec-bot]

Summary

• add DFlash2 draft-model support using the official grouped dynamic convolution and candidate-selector design
• dispatch DFlash2 checkpoints through the existing DFLASH runtime
• connect the draft q distribution to stochastic verification with min(1, p / q) acceptance and residual relu(p - q) sampling
• support GQA and MLA DFlash2 checkpoints, including Kimi-K3 target-hidden capture and MLA context-cache writes
• keep Kimi-K3 draft MLA KV in the unified full-retention cache group while carrying per-layer sliding-window metadata on the compute attention objects
• keep CUDA Graph compatibility and fail fast for unsupported random-sampling backends
• add focused tests and user-facing configuration documentation

Validation

• DFlash2 and Kimi-K3 focused tests: 17 passed
• Kimi-K3 BF16 target plus DFlash2 MLA draft loaded at TP8 and captured CUDA Graph buckets [1, 2, 3, 4]
• GSM8K, 128 samples, concurrency 1, temperature 1.0, top-p 0.95, top-k 20: mean acceptance length 5.271, median 5.403, proposal acceptance 61.01%
• GSM8K boxed-or-last-number accuracy: 114/128 (89.06%)
• pre-commit run --all-files passed

WIP

• broader topology qualification and performance tuning will follow