feat: add non-greedy MTP sampling kernels

[cl-vv-h]

Motivation

NPU MTP verification currently supports greedy verification, but non-greedy
requests do not have a corresponding probability-sampling path. This can make
temperature, top_k, and top_p ineffective during speculative
verification.

With GLM-4.7-Flash on GPQA, enabling non-greedy MTP previously reduced the
observed accuracy from approximately ~70% to ~50%.

Modifications

• Add sequential top-k then top-p probability renormalization.
• Add target-only non-greedy speculative sampling for general EAGLE trees.
• Add classic rejection sampling for tree_topk=1 linear chains:
  • accept draft token x with min(1, p(x) / q(x));
  • sample from relu(p - q) after the first rejection;
  • sample from the target distribution when all drafts are accepted.
• Use block-wise vocabulary reduction and inverse-CDF sampling for large
  vocabularies.
• Export the new sampling APIs from sgl_kernel_npu.sample.
• Keep the rejection loop-carried cur_prob_row explicitly typed as int64
  for Triton SSA type consistency.

Supported Scope

• Target-only sampling supports general EAGLE trees with tree_topk >= 1.
• Classic rejection sampling supports tree_topk=1 linear chains.
• Probability tensors use FP32 at the kernel boundary.
• Existing greedy sampling code and behavior are unchanged.
• Unsupported shapes, dtypes, devices, and topologies fail explicitly.

Tests

export PYTHONPATH="$PWD/python/sgl_kernel_npu:${PYTHONPATH:-}"
python -m pytest -q -s \
  tests/python/sgl_kernel_npu/test_speculative_probability.py \
  tests/python/sgl_kernel_npu/test_tree_speculative_sampling_target_only.py \
  tests/python/sgl_kernel_npu/test_chain_speculative_sampling.py

The tests cover probability renormalization, general-tree target-only
sampling, classic rejection chains, early rejection, all-accepted requests,
acceptance thresholds, and vocabularies up to 151552 tokens.

Accuracy

Model: GLM-4.7-Flash
Dataset: GPQA_Diamond

Configuration	Before	After
NPU MTP target-only	~50%	~70%
NPU MTP rejection (tree_topk=1)	~50%	~70%

Performance

Mode	Result
Target-only	No measurable performance impact
Classic rejection	No measurable performance impact

Checklist

• Code formatting checks completed.
• Unit tests completed.
• Accuracy check completed.
• Performance check completed.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.