[AMD][DI][CI] Seed the draft proposal distribution on the PD decode path

[Lzy17]

GLM-5.2 EP16 with MTP dies about a minute into serving:

File "speculative/eagle_worker_v2.py", line 932, in draft_forward
  torch.stack(draft_probs_list, dim=1)
TypeError: expected Tensor as element 0 in argument 0, but got NoneType

Cause

build_eagle_disagg_draft_input is the only constructor of EagleDraftInput
on the PD decode path. The prefill worker sends topk_p and topk_index but
no proposal distribution, so it leaves draft_probs unset, and the eager draft
loop seeds its list with that field:

# eagle_worker_v2.py:719
if get_spec().speculative_use_rejection_sampling:
    draft_probs_list: List[torch.Tensor] = [spec_info.draft_probs]

The list is stacked at the end of the loop, so a None seed is a guaranteed
TypeError as soon as the first draft step completes.

The graph path never hit this because EAGLEDraftCudaGraphRunner allocates a
zeroed draft_probs buffer under the same condition
(eagle_draft_cuda_graph_runner.py:197). The eager path had no equivalent.
Every MTP test that passes today captures graphs; this recipe sets
--disable-cuda-graph.

Scope: this is not AMD-only

The tag is on the PR because that is where the nightly caught it, but the
defect is in common code and reachable on any backend:

• speculative_use_rejection_sampling is a documented user-facing flag
  (fields/spec.py:116), default False. Any CUDA user who sets it explicitly
  and runs PD + EAGLE without CUDA graphs hits the identical crash. ROCm only
  differs in that _should_auto_enable_hip_rejection_sampling turns it on by
  default, so AMD reaches the path without opting in — which is why the nightly
  found it first.
• [PD] Keep EAGLE DP graph and token metadata consistent #32196 (merged) makes the eager path easier to reach for everyone: when
  topk == 1 and a DSA seed is required but dsa_topk_indices is None, the
  batch is now marked cuda_graph_compatible=False and falls back to eager.
  That is exactly the path with the missing seed.

Fix

Allocate zeros under the same condition the graph runner already uses, so the
two paths agree. Zeros are the correct seed rather than a placeholder: the
sampler rejects q == 0 and resamples that position from the target
distribution, which is the behaviour the zeroed graph buffer has always
produced.

Rebased onto #32196; the two changes are independent additions to the same
function.

Verification

MI355X nightly, ROCm 10, GLM-5.2 EP16 + MTP, PD 1P1D with
--disable-cuda-graph: crashes on main, serves to completion with this
patch, GSM8K 0.949.

---

CI States

Latest PR Test (Base): ✅ Run #36804310116
Latest PR Test (Extra): ❌ Run #36804309955
Latest PR Test (AMD ROCm 10): ❌ Run #36804310105

[Lzy17]

Confirmed on MI355X, ROCm 10 (v0.5.20-rocm10-mi35x-20260923), GLM-5.2 FP4
2P1D EP16 + MTP:

	before	after
GSM8K	crashed before scoring	0.949
concurrency sweep	0/7	7/7
draft_probs_list TypeError	yes	none

One note on why zeros are the right value rather than a workaround: the graph
path's buffer is also permanently zeros here. replay only copies into it when
spec_info.draft_probs is not None, which on the PD path it never is, and
nothing else writes it. So this makes the eager path match the graph path
exactly.