…l only

The fused Triton kernel introduced in PR sgl-project#23620 (commit adc5932) is
correct enough for non-speculative target-model decode (its original
target, MiniMax-M2.5) but its bf16->fp8 implicit cast through tl.store
does not match PyTorch .to(torch.float8_e4m3fn) bit-exactly. PyTorch
casts with round-to-nearest-even + saturation; the Triton path on
ROCm/HIP rounds differently and may not saturate, even when the
per-tensor k_scale / v_scale are 1.0 (verified for Kimi-K2.5 Quark
MXFP4 with kv_cache_dtype=fp8 by direct probe).

Non-speculative inference tolerates this small numerical drift, but
EAGLE3 draft decode reads back its own freshly written K/V cache on
every subsequent draft step, so any drift in the draft cache compounds
across draft steps and collapses the accept length:

  Kimi-K2.5-MXFP4 + EAGLE3 (8xMI300, in/out 8192/1024, conc 4):
    pr-23146 baseline               : accept=3.26  out=675 tok/s
    + seqused_k fix (2bee3c3)     : accept=3.46  out=706 tok/s
    + this commit (target-only gate): accept=3.97  out=807 tok/s
  pr-23461 baseline reference       : accept=3.97  out=798 tok/s

Restrict the fast path to target-model backends by checking
model_runner.is_draft_worker. The SWA path is unchanged (it already
works because SWA models did not exercise the corrupted draft cache).
The Triton kernel itself can be revisited later to match PyTorch fp8
cast semantics; until then, draft model writes route through the
legacy MHATokenToKVPool.set_kv_buffer path.

…ache write in aiter decode path (sgl-project#23620)