refactor: unify DSpark heads and load the DeepSeek-V4 confidence head.

[weizhehuang0827]

Description

Two related DSpark head cleanups plus a fix that were split out of the DSpark
adaptive-decode work:

1. Relocate the confidence head. Move dspark_confidence_head.h from
   models/llm/npu/ to models/llm/ (namespace xllm), next to
   dspark_markov_head.h. It is plain torch code with no NPU dependency, so it
   does not belong under npu/.

2. De-duplicate DSparkMarkovHead. qwen3_dspark.h carried its own inline
   copy of the Markov head that had diverged from the shared
   models/llm/dspark_markov_head.h. Both the Qwen3 and DeepSeek-V4 drafts now
   use the shared head (which gains the markov_embed() helper the confidence
   head reuses); the Qwen3 loader strips the markov_head. prefix so the shared
   unprefixed keys match.

3. Load the DeepSeek-V4 DSpark confidence head. The DeepSeek-V4-Flash
   checkpoint ships mtp.<last>.confidence_head.proj.weight (with-markov,
   bias-less), but DeepseekV4DSpark never loaded it, so adaptive pruning
   silently fell back to sampler proposal probabilities instead of the trained
   confidence. DSparkConfidenceHead now treats the bias as optional, and the
   DeepSeek-V4 model loads and exposes the head through the same interface as
   Qwen3.

The confidence head is only stood up when adaptive speculative decode is enabled
(SpeculativeConfig::enable_adaptive_speculative_decode()); under static
decoding its weights are never loaded, since only the adaptive controller
consumes them.

Test

• Builds cleanly (including the dspark_confidence_head_test numerical-parity
  unit test, which was updated for the shared Markov head's constructor and
  unprefixed keys).
• Runtime on DeepSeek-V4-Flash, 8 NPUs, dp_size=2, ShareGPT:
  • static (adaptive off): confidence head is not loaded (no verify failure)
    and serving is unaffected.
  • adaptive (adaptive on): the confidence head loads and is consumed by the
    controller; the run serves all requests with no crash or HCCL hang.