[pytorch/pytorch] torch.compile crashes on user-defined Triton kernel referencing triton.language.dtype globals

### ROOT CAUSE The crash occurs because `torch.compile` does not properly handle user-defined Triton kernels that reference `triton.language.dtype` globals. The compiler fails to resolve these globals during the compilation process, leading to a crash. This is due to the interaction between PyTorch's compiler and the Triton JIT system, where certain global references are not recognized. ### CODE FIX Modify the kernel to avoid using the `dtype` global from `triton.language` by passing the dtype as an argument instead. This ensures compatibility with `torch.compile`. ```python import triton import triton.language as tl @triton.jit def kernel(x, dtype: tl.dtype): # Use the passed `dtype` argument instead of referencing the global pass ``` This change ensures the kernel's dependencies are explicitly provided, making it compatible with `torch.compile`.