Update torch pin to Sep 06, 2025 (pytorch#14129)

Summary:

Branch cut for PT/PT is Sep 08, so let's bring it forward just a few
days before the branch cut.

Once we cut the 1.0 branch cut in ET, we will pin to 2.9 which is
basically equivalent of Sep 08

Author: mergennachin

.ci/docker/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-e7152ff8a6a929a0db7f3f4a72a5b6d471769cd3
+4d4abec80f03cd8fdefe1d9cb3a60d3690cd777e

install_requirements.py

@@ -71,7 +71,7 @@ def python_is_compatible():
 #
 # NOTE: If you're changing, make the corresponding change in .ci/docker/ci_commit_pins/pytorch.txt
 # by picking the hash from the same date in https://hud.pytorch.org/hud/pytorch/pytorch/nightly/
-NIGHTLY_VERSION = "dev20250811"
+NIGHTLY_VERSION = "dev20250906"
 
 
 def install_requirements(use_pytorch_nightly):

runtime/core/portable_type/c10/torch/headeronly/macros/Macros.h

@@ -259,7 +259,8 @@ using namespace c10::xpu;
 // to resolve potential warnings.
 #if __CUDA_ARCH__ == 750
 constexpr uint32_t CUDA_MAX_THREADS_PER_SM = 1024;
-#elif __CUDA_ARCH__ == 860 || __CUDA_ARCH__ == 870 || __CUDA_ARCH__ == 890
+#elif __CUDA_ARCH__ == 860 || __CUDA_ARCH__ == 870 || __CUDA_ARCH__ == 890 || \
+    __CUDA_ARCH__ == 1200
 constexpr uint32_t CUDA_MAX_THREADS_PER_SM = 1536;
 #else
 constexpr uint32_t CUDA_MAX_THREADS_PER_SM = 2048;