[None][fix] Share Eagle3 draft KV layers in aggregated serving, not only under attention DP

[zheyuf]

Description

One-model speculative decoding places the drafter's KV either in the target
manager's blocks (shared) or in a separate draft KV cache manager.
Today that choice is gated on attention DP:

if self._mapping.enable_attention_dp and getattr(cls, 'supports_shared_draft_layers', True):
    return False  # shared
return should_use_separate_draft_kv_cache(self._speculative_config)

The gate reflects the layout existing deployments were validated with, not a
technical requirement. The consequence is that aggregated serving with
attention DP off takes the separate manager for every model using
Eagle3-one-model / MTP one-model.

Why that costs acceptance

When the target reuses a prefix it does not recompute those tokens. The
drafter therefore never gets a chance to compute its own KV for them either,
and it enters generation blind over the reused span — its drafts are rejected
until it regrows the state. Sharing the layers removes the failure mode
structurally: one set of blocks, one lifetime, one reuse decision.

Measurement

Controlled warm/cold probe: 32 x 1684-token prompts sent twice in the same
process. Round 2 hits the prefix cache, so the two rounds have identical
prompts and identical work and differ only in where the drafter's prefix KV
comes from. MiniMax-M3 NVFP4 target, TP4/EP4, Eagle3 draft_len 3, greedy.

arm	cold AL	warm AL	delta
separate manager, 4-KV-head Eagle3 head	3.458	3.195	−0.263
separate manager, 64-KV-head Eagle3 head	3.555	3.282	−0.273
shared layers, 4-KV-head head	3.553	3.617	+0.064
shared layers, 64-KV-head head	3.603	3.596	−0.007
reuse disabled (control), both heads	—	—	±0.02

Two Eagle3 heads that differ only in num_key_value_heads (64 vs 4) lose the
same ~7.5%, so this is a property of the separate-manager path rather than of
a particular drafter. The reuse-disabled control rules out a systematic
round-2 effect.

Scope

• Models keep the ability to opt out through supports_shared_draft_layers;
  MiniMax-M3 uses that to keep the separate manager for its coalesced
  index-K pool.
• Disaggregated serving already forced sharing (the nvbugs/5807902 WAR), so
  this only changes the aggregated non-DP path.
• Sharing has been exercised in CI on the attention-DP path since it was
  introduced; this widens which configurations reach it.

Draft — what still needs doing before this leaves draft

• Confirm on a second model family (a Llama/Qwen Eagle3 or an MTP model)
  that aggregated non-DP acceptance improves and nothing regresses.
• Memory review: the draft layers now allocate from the target's pool
  instead of their own budget split; check _split_kv_cache_budget_for_draft
  is not left double-counting for the shared path.
• CI: an aggregated non-DP acceptance point, so the regression cannot come
  back silently.

PR Checklist

• PR title follows the [JIRA/NVBUG/None][type] format
• Commits are signed off (DCO)
• CI (/bot run)