cuda: tune fixed-size softmax launch

[john-sp]

• Reduce the optimized 64-element softmax launch from 256 to 128 threads.
• Compute one exact reciprocal and reuse it for the two outputs handled by each thread.

Performance:

On RTX 3090 Ti at batch 84:

• Current paired WSL Nsight trace: 18.372 us -> 18.454 us median, effectively neutral.
• Earlier cleaner isolated traces: 19.828 us -> 19.409 us, approximately 2.1% faster.

AI Acknowledgement: Written by Codex

[Copilot]

Pull request overview

This PR tunes the CUDA-optimized softmax path for C == 64 by adjusting the kernel launch configuration and reducing redundant arithmetic in the normalization step.

Changes:

• Reduce the softmax_opt_64_kernel block size from 256 to 128 threads (4 warps / 4 rows per block).
• Compute a single reciprocal of the warp-wide sum (inv_sum) and reuse it for both outputs per thread.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.