uv: Deduplicate all files in the wheel cache

The patch replaces per-file 64 KiB buffer allocations with a single reusable buffer for each wheel, cutting buffer creations from 11,120 to one for the PyTorch wheel while maintaining a 64 KiB size. Benchmarks on Linux show cold-install times decreasing 2.6%–9.5% for individual packages and about 7% for a 14-package environment with concurrency 4.