perf(K2.5): optimize small kernels in EAGEL3 drafter loop

[syuoni]

Summary

Per-call overhead reductions in the EAGLE3 drafter loop and surrounding metadata prep.

Changes

• compute_out_cache_loc uniform variant — added compute_out_cache_loc_uniform
  for the drafter's multi-step decode where every request has input_length=1. Skips
  the per-call torch.cumsum + torch.full host-side work and the kernel's GMEM
  reads of input_lengths_ptr/cumsum_lengths_ptr (Triton specializes on
  None-pointer at JIT time).
• req_pool_indices_buf is now int64 — eliminates ~12 implicit int32→int64
  unrolled_elementwise cast kernels (~1.6 µs each) along the per-iteration
  metadata prep path. Pairs with switching the valid_cache_lengths[idx] fancy
  index to valid_cache_lengths.index_select(0, idx) so int32 indices are still
  accepted natively where they remain (e.g. in index_add_).
• Persistent drafter buffers — draft_seq_lens_buf, draft_out_cache_loc_buf,
  draft_input_lengths_buf, and last_index_offsets_buf (= arange(max_bs) * spec_num_tokens - 1) are hoisted to Eagle.__init__ to avoid per-call alloc +
  init. last_index_offsets is now plumbed via ForwardContext to
  LogitsProcessor for the padded-static-len last-token-select path.
• Eagle.draft() cleanup — fused cache_start + 1 into
  torch.add(..., out=draft_seq_lens); replaced the post-draft() torch.cat
  with direct writes into a shared next_tokens[bs, spec_num_steps+1] buffer;
  skip the last-iter positions.add_(1) / draft_seq_lens.add_(1) since they're
  not consumed; removed the dead logits.shape[0] != bs fallback branch.
• Persistent drafter buffers — draft_seq_lens_buf, draft_out_cache_loc_buf,
  draft_input_lengths_buf, and last_index_offsets_buf are hoisted to
  Eagle.__init__ to avoid per-call alloc + init. The last one
  (= torch.arange(max_bs) * spec_num_tokens - 1, int64) replaces two
  per-call torch.arange(bs, ...) * spec_num_tokens patterns — one in the
  drafter's last-verified-id selection, one in LogitsProcessor's padded-static-len
  last-token-select. The precomputed buffer is sliced to [:bs] and plumbed via
  ForwardContext.last_index_offsets so LogitsProcessor can skip the
  arange + mul + sub triplet. Net: pre-step-0 last-token-select drops from
  6 kernels (arange + mul + sub + 2 gathers) to 3 (1 add + 2 gathers).

[LorrinWWW]

lgtm