[Spec Decoding] Integrate DFlash into speculative decoding pipeline

[aaronzhfeng]

Description

Wire DFlash block-diffusion speculative decoding into the existing TPU inference pipeline. The DFlash model and proposer were added in #1868; this PR connects them to the runner, KV cache manager, and speculative decoding manager so DFlash can be used end-to-end.

No changes to existing Eagle3 or ngram code paths: DFlash gets its own propose_dflash_draft_token_ids method and a separate elif "dflash" dispatch branch.

Modified files:

• tpu_inference/models/common/model_loader.py -- register DFlashDraftModel in model registry
• tpu_inference/models/jax/qwen3.py -- collect aux_hidden_states from target layers during forward pass (needed by DFlash proposer to inject target context)
• tpu_inference/runner/tpu_runner.py -- add DFlashProposer initialization for method="dflash"
• tpu_inference/runner/speculative_decoding_manager.py -- add dflash method dispatch and propose_dflash_draft_token_ids (uses accepted_attn_metadata with correct seq_lens for drafter)
• tpu_inference/runner/kv_cache_manager.py -- extend draft KV cache allocation to cover dflash, read num_hidden_layers from config instead of hardcoding 1

Usage (after both #1868 and this PR):

args['speculative_config'] = {
    'model': 'z-lab/Qwen3-4B-DFlash-b16',
    'num_speculative_tokens': 5,
    'method': 'dflash',
    'draft_tensor_parallel_size': 1,
}

Tests

E2e tests are in a follow-up PR.

Checklist

• I have performed a self-review of my code.
• I have necessary comments in my code, particularly in hard-to-understand areas.
• I have made or will make corresponding changes to any relevant documentation.