Allow FP16 math in flash attention (#24953)
### Description
This change restores back fp16 math based FlashAttention.
### Motivation and Context
Earlier we noticed quality issues with deepseek-r1 attributed to
overflow of qk computation when performing math in fp16 precision.
https://github.com/microsoft/onnxruntime/pull/24723, addressed it by
promoting math to fp32 to avoid the precision issue.
However the topic remained that, these models are trained with FP8
precision how is it that inferencing runs into precision issues with
FP16 math? using FP32 math also resulted in slight performance
degradation.
In this follow up investigation, one issue identified is that we
multiply scale for gqa quite late in the computation. Scale is 0.088 for
deepseek-r1. Multiplying scale upfront seems to prevent the overflow
issues.
For now only the prefill shaders are updated to use this approach.
Pending feedback on impact across models, the generation shader can also
be restored to FP16 math.
---------
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
