PyTorch migrates upsampling ops to Metal kernels, fixes MPS bug

PyTorch's latest commit to its MPS (Metal Performance Shaders) backend replaces the MPSGraph implementation of 1D/2D nearest and nearest-exact forward upsampling with custom Metal kernels. This migration fixes a latent correctness bug in the MPSGraph path and unifies the host-side dispatch logic: a single templated launcher now drives every Metal upsample kernel, while a new UpsampleParams struct supplies geometry for all ranks. The change collapses previously duplicated per-rank dispatch overloads, streamlining the codebase. The linear, bilinear, bicubic, and 3D forward paths were already running on Metal, so this completes the transition for all forward upsampling operations.

Interestingly, the backward pass remains on MPSGraph. The developers intentionally kept it there because MPSGraph's fused gradient kernel performs several times faster than a naive Metal gather implementation. This decision preserves performance where it matters most. The pull request (#186989) was approved by PyTorch maintainers and co-authored by Claude, Anthropic's AI coding assistant. For developers using PyTorch on Apple Silicon, this means more reliable and potentially faster upsampling during model inference, while training gradients still benefit from MPSGraph's optimised backward kernels.