PyTorch slashes Docker build times with shared registry cache

PyTorch's CI has long relied on `docker-builds` running on a remote BuildKit fleet (OSDC). Each build previously passed `--no-cache`, forcing a complete rebuild of every Docker layer — a slow and resource-heavy process. The ephemeral pods also couldn't share their private caches, compounding the inefficiency. Now, a new pull request (#186297) replaces that approach with a registry-based caching system. For remote builds, the CI now uses `--cache-from` and `--cache-to` with `type=registry`, writing cache layers to a per-image tag (e.g., `<image>-buildcache`) in the same ECR repository. The cache is configured with `mode=max` (stores all layers) and `image-manifest=true,oci-mediatypes=true` to ensure compatibility.

This change means unchanged stages are reused across CI runs, dramatically cutting build times and reducing the load on the shared BuildKit fleet. A companion PR sets up a lifecycle rule to delete these cache tags after 7 days to prevent storage bloat. Developers testing in CI will see faster iterations, and the PyTorch maintainers benefit from reduced infrastructure costs. This is a behind-the-scenes optimization that makes PyTorch's development pipeline more efficient.