trunk/f7c9c9f812f57c14ade218e76c6a3003594d7128: [MPS] Add API to load pre-compiled metallib (#177276)

The PyTorch team has merged a key technical update to its Metal Performance Shaders (MPS) backend, introducing a new low-level API: `torch._C._mps_loadMetallib(bytes)`. This function allows developers to directly load pre-compiled `.metallib` binary blobs into PyTorch's execution runtime on macOS. It works by calling Apple's Metal `newLibraryWithData:` method, returning a library object that provides direct access to GPU kernel functions. This is a companion to the existing `_mps_compileShader(source)` API but crucially skips the expensive runtime compilation step.

This change is not a standalone feature but a foundational building block required for the upcoming Triton compiler backend for Apple MPS. The Triton compiler, used for writing highly optimized GPU kernels in Python, can now generate `.metallib` files from LLVM Intermediate Representation (IR) at compile time—long before a user runs their PyTorch model. By loading these pre-built binaries, frameworks can eliminate the just-in-time (JIT) compilation delay that occurs the first time a new neural network layer or operation is executed on an Apple GPU, leading to dramatically faster model load times and more responsive development cycles.

The pull request (#177276) was approved by core maintainers and links directly to the ongoing integration work in the Triton-Lang repository (triton-lang/triton#9701). For developers and researchers using Macs with M1, M2, or M3 chips for machine learning, this backend optimization means that complex models leveraging custom Triton kernels will start up much faster, making local AI development and experimentation on Apple hardware more efficient and competitive with other platforms.