Add configurable GQA and MLA sliding-window draft layers

[maocheng23]

Dependency

This is a stacked draft on #758. The first commit is the generic MLA support from that PR; after #758 merges, this PR will contain only the SWA configuration delta.

Summary

• make context layout and projection layout independent for every DFlash-family draft layer
• keep dflash_config.attention_mode as the backward-compatible uniform GQA/MHA/MLA shorthand
• add dflash_config.attention_modes for per-layer GQA/MHA/MLA selection
• support SWA-GQA, SWA-MLA, full-context GQA/MLA, and mixed stacks without a model-specific architecture
• validate ambiguous, invalid, and wrong-length configurations early
• document the paired layer_types / attention_modes contract

layer_types continues to select full_attention or sliding_attention; the projection list is paired with it by layer index. The GQA and MLA parameterizations follow the standard layouts used by TorchSpec, while SpecForge's existing DFlash mask and objective interfaces remain unchanged.

Example

{
  "num_hidden_layers": 4,
  "layer_types": [
    "sliding_attention",
    "sliding_attention",
    "full_attention",
    "full_attention"
  ],
  "use_sliding_window": true,
  "sliding_window": 2048,
  "dflash_config": {
    "attention_modes": ["gqa", "mla", "gqa", "mla"]
  }
}

Validation

• all pre-commit hooks passed
• CPU targeted regression: 34 passed, 2 skipped, 19 subtests passed
• H200 mixed SWA-GQA/SWA-MLA FlexAttention forward/backward: 6 passed, 7 subtests passed
• broader modeling/algorithm/runtime/training regression: 93 passed, 5 skipped, 54 subtests passed
• tested with Transformers 5.8.1 and PyTorch 2.11.0+cu130