fix(kernel): make Triton top-k/top-p sampling exact

[taking-lying-flat]

Summary

• incorporate the exact single-launch fp32 histogram search from fix(kernel): exact single-launch triton top-k/top-p sampling #345
• clip top-k and top-p boundary ties deterministically by token id, including combined top-k → top-p filtering
• retry the same exact kernel with one CTA per row when cooperative launch is unavailable
• scope cooperative-launch capability by device and operation, and propagate unrelated Triton failures
• remove the older approximate multi-launch fallback and redundant workspace/target conversions

This keeps the final PR to one production file; no test files are included.

Validation

• local-only CUDA regression suite: 9 passed
• randomized 64 × 4096 comparison against stable PyTorch top-k/top-p references, including sampled-token support checks
• Ruff, Python byte-compilation, and diff whitespace checks
• RTX A1000 benchmark against fix(kernel): exact single-launch triton top-k/top-p sampling #345: comparable latency, with no additional full-vocabulary tie pass

Credit

This revision builds directly on #345. Thank you, @jason-fxz, for taking the original correctness report seriously, replacing the fragile estimator with a much stronger exact design, and carefully reviewing #329. The core single-launch histogram approach here is your work; these follow-up changes tighten its boundary semantics and fallback behavior. I genuinely appreciate the time and thought you put into it.