Update PyTorch pin + apply flex attn 143553 patch offline (#4592)

Flex attn causal mask:
https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/15920836126/job/44907063106
(same perf)

The offline patch will allow updating to other versions of PyTorch (not
necessarily to the one that may suddenly start to be used in the pull
request). Now I need this to reduce the amount of things that can
contribute to the performance regression that is observed in
#4553


For the problematic inductor test - I have opened a separate issue for
now: #4593

Signed-off-by: Anatoly Myachev <[email protected]>

Author: anmyachev

.github/pins/pytorch.txt

@@ -1 +1 @@
-29e6033ff38f0deb8dde1146a1612e64ed00e3b7
+c2d1b225e62bdf3adbba91891f53bc60315adaac

scripts/patch-pytorch.sh

@@ -35,6 +35,6 @@ apply_patch() {
 echo "Applying PyTorch patches in $REPO_ROOT"
 
 # put your patch applies here
-apply_patch https://github.com/pytorch/pytorch/pull/143553.diff
+apply_patch ./patch/flex_attn_143553.patch
 apply_patch pytorch_fp64.patch
 apply_patch ./patch/Patch_torch_flex_attention_for_autotune_in_benchmark.patch