[webgpu ] Optimize flash attention for Nvidia (#25777)
In the flash attention algorithm, each thread in a subgroup needs to
access the same range (0-15) of data in workgroup memory `q_tile` and
`v_tile`. If we use `subgroupShuffle`, there will be bank conflicts for
`var k_local = k_tile[capped_sg_id][i];` since the sg_size is 32 and
thread16~thread31 are accessing the same bank address.
To avoid the bank conflicts, we can directly access the same address in
workgroup memory by all threads which is a broadcast and well optimized
in the NV GPUs. See ~10% improvement for phi4 prefill (1K) in NV RTX
2000 Ada. And as the input gets longer(total_sequence_length), the
optimization effect gets better (~12% for 2K).
Before
```
Batch size: 1, prompt tokens: 1000, tokens to generate: 128
Prompt processing (time to first token):
avg (us): 2.0991e+06
avg (tokens/s): 476.394
p50 (us): 2.08457e+06
stddev (us): 36140.3
n: 5 * 1000 token(s)
Token generation:
avg (us): 25477.8
avg (tokens/s): 39.2498
p50 (us): 25028.2
stddev (us): 4841.89
n: 635 * 1 token(s)
```
After
```
Batch size: 1, prompt tokens: 1000, tokens to generate: 128
Prompt processing (time to first token):
avg (us): 1.91138e+06
avg (tokens/s): 523.183
p50 (us): 1.92379e+06
stddev (us): 44768
n: 5 * 1000 token(s)
Token generation:
avg (us): 25237.2
avg (tokens/s): 39.624
p50 (us): 24860.9
stddev (us): 4874.52
n: 635 * 1 token(s)
```
