Optimize topksoftmax: top-K-only softmax + 32B vector loads

[CuiCu-618]

Motivation

Technical Details

• Fuse k-select with softmax so we compute exp() only for top-K
• Widen vector loads to 32B

Test Plan

Test Result

MI300A

N=64, K=5-8: ~1.12× - 1.53x kernel speedup;
N=256, K=5-8: ~0.97× - 1.13x kernel speedup;
N=512, K=5-8: ~0.98x - 1.76x kernel speedup;

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.