fix(engine): preserve greedy sampling in mixed batches

[taking-lying-flat]

A greedy request with fixed logits [1.0, 1.0, 0.0] returns token 0 alone, but can return token 1 when batched with a random request: the mixed path approximates greedy with temperature 1e-6, leaving the tied maxima equally likely. Also, temperature=0, top_p=0.9 unnecessarily enters probability sampling.

Keep a per-request greedy_mask and explicitly select argmax for greedy rows. Treat temperature <= 0 or top_k == 1 as greedy regardless of top_p, and use neutral backend parameters for those rows. The all-greedy fast path is preserved. Mixed batches still run probability sampling across all rows; skipping that computation is outside this correctness fix.

Related: #307. This fixes sampling-policy differences for identical logits; it does not establish or resolve the cause of the serial model-level divergence reported there.

Validation:

• Ran local deterministic CPU checks covering tied maxima, standalone versus mixed batches, FP32/FP16/BF16, and all four top-k/top-p modes. Before the fix, the mixed regression returned [1, 2, 2] instead of [0, 2, 1], and zero-temperature requests with top_p < 1 entered probability sampling. The validation file is retained locally and is not included in this PR.
• On upstream base e0886cc plus this patch: 73 passed, 4 skipped, including real Triton CUDA sampling. The four FlashInfer cases were skipped because FlashInfer is not installed. git diff --check passed.
• Hardware: NVIDIA RTX A1000 Laptop GPU; Intel Core i7-12800HX; NVIDIA driver 595.84; PyTorch 2.11.0+cu130. Checkpoint: N/A, tests use fixed logits without loading a model.

Exact local validation command, from the repository root, using the untracked test file:

env PYTHONPATH=python /tmp/freetoken-glm-shared-venv/bin/python -m pytest tests/engine/test_sample.py -q -rs