[shihaobai]

GLM NextN draft Prefill can discard the draft output after its cache state is committed. This change keeps the shared BaseModel flow unchanged and moves the optimization into the GLM MTP model:

• The draft-specific Prefill layer writes MLA KV, indexer K-pools, and the raw tail directly, then returns a zero-row placeholder.
• The draft post layer converts that placeholder into an empty output, skipping the unused Q-B projection, sparse attention, FFN, final hidden collection, and logits.
• EAGLE uses the cache-only path for all draft steps. Vanilla With-Att uses it for a single chained draft; multi-level Vanilla still runs the intermediate outputs needed by the next draft.
• The same writer handles BF16/FP8 layouts, TP/TPSP, DP prefill balance, mixed prefill, and overlap execution.

Cache-only writes are byte-exact with the full writer across HOLD rows, short and 8K prompts, long-prefix/pool-boundary requests, and mixed request layouts.