Verify that the fused kernel returns logits matching a PyTorch matmul reference.

Verify that greedy_sampling=True returns the argmax token for each sequence.

Verify that top-k and top-p filtering restricts samples to the expected tokens.

Test top-k + top-p at V=151936 with real Qwen3-0.6b weights. At large vocab sizes, tl.dot and cuBLAS produce different logits due to bf16 acc