[Feature] Add MLA draft model support for Eagle3 training

[cicirori]

Summary

• Add DeepSeek MLA (Multi-head Latent Attention) draft model for Eagle3 online training
• Support both SDPA and flex_attention backends
• Enable using MLA-based draft models (DeepSeek-V2/V3 style) with any target model
• Reference config based on nvidia/Kimi-K2.5-Thinking-Eagle3
• Fix convert_to_hf.py: lm_head pruning bug, draft_vocab_size validation, train-time pruning detection

Changes

• torchspec/models/draft/deepseek_eagle.py — DeepSeekMLAAttention, DeepSeekMLAFlexAttention, DeepSeekDecoderLayer, DeepSeekForCausalLMEagle3
• torchspec/models/draft/auto.py — Register DeepseekV3Config → DeepSeekForCausalLMEagle3 dispatch
• configs/draft_models/qwen3_8b_eagle3_mla.json — Qwen3-8B target + MLA draft config (full vocab, no train-time pruning)
• configs/sglang_qwen3_8b_mla_draft.yaml — E2E training config (flex_attention)
• tests/test_deepseek_eagle.py — 12 unit tests (shapes, gradients, config dispatch, softmax scale, TTT loop, SDPA vs flex consistency)
• tools/convert_to_hf.py:
  • Fix lm_head trimming bug (arange + d2t → direct d2t indexing)
  • Error when draft_vocab_size != vocab_size without --prune-vocab (model needs t2d/d2t)
  • Error when lm_head already pruned (train-time pruning incompatible with post-training re-prune)
• E2E training loop fixes for single-node runs

Test plan

• Unit tests: 12/12 passed (CPU + CUDA, SDPA + flex_attention)
• E2E: ./examples/qwen3-8b-single-node/run.sh configs/sglang_qwen3_8b_mla_draft.yaml — 500 steps on B200
• convert_to_hf.py without --prune-vocab → full vocab model, from_pretrained OK
• convert_to_hf.py with --prune-vocab --draft-vocab-size 32000 → pruned model with t2d/d2t, from_pretrained OK
• pre-commit (ruff + ruff-format + isort) all pass

E2E training result (500 steps, Qwen3-8B + MLA draft, flex_attention, B200)

$ ./examples/qwen3-8b-single-node/run.sh configs/sglang_qwen3_8b_mla_draft.yaml

Training: 100% | 500/500 [01:45<00:00, 4.75step/s, loss=1.274, acc=0.865, acc_len=4.15, thru=6.5, I=145.7, T=12.6, wait=0.0s, pool=0, epoch=1/1]

HF conversion verification

# Without --prune-vocab (full vocab)
$ python tools/convert_to_hf.py --input-dir outputs/.../iter_0000501 --config configs/draft_models/qwen3_8b_eagle3_mla.json -f
vocab=151936/151936, lm_head=[151936, 4096], has_vocab_pruning=False, params=1,493,714,944

# With --prune-vocab (post-training pruning)
$ python tools/convert_to_hf.py ... --prune-vocab --draft-vocab-size 32000 --dataset-path examples/data/sample_conversations.jsonl --tokenizer Qwen/Qwen3-8B --chat-template qwen --prompt-key conversations
vocab=32000/151936, lm_head=[32000, 4096], has_vocab_pruning=True, params=1,002,457,088

# Both load correctly via from_pretrained

FSDP2 compatibility note

MLA's bottleneck structure (down_proj → RMSNorm → up_proj) is compatible with TorchSpec's per-Linear FSDP2 sharding. FSDP2 shards parameters only — weights are all-gathered before forward, so activations (including the compressed latent and torch.split on kv_a_proj output) are always full-size during computation. RMSNorm layers are not nn.Linear and remain replicated. This is the same sharding pattern used by the existing Llama attention, verified end-to-end on B200 (500 steps, loss converges to 1.274).