binary add become slower than 2.6

[jianyizh]

Run this with unitrace -d --conditional-collection python test_add.py
Current pytorch main gets 2342624ns on average, pytorch 2.6 gets 951728ns on max 1550.

I noticed that Register File Size Per Thread increased from 128 to 256

import torch
import os
 
input_tensor_16 = torch.randn(16,12,512,512, dtype=torch.bfloat16, device="xpu")
input_tensor_32 = torch.randn(16,1,1,512, device="xpu")
 
for _ in range(10):
    _ = input_tensor_16 + input_tensor_32
 
torch.xpu.synchronize()
os.environ['PTI_ENABLE_COLLECTION']="1"
for _ in range(10):
    _ = input_tensor_16 + input_tensor_32
torch.xpu.synchronize()

[jianyizh]

pytorch/pytorch@382fbcc#diff-d8b83827d6e6c37c12a967f9c0b9cf2c93da12e9372f95302e0051b6f2526d3e This pytorch commit add fp8, leads to igc auto grf choose 256 instead of 128 register files for cast, so it will be slower for large shapes

[chuanqi129]

PR for is ready for review, need to decide whether to cherry-pick to release 2.7