Persistent Kernel Cache

There are a few issues enabling this currently:

1. It turns on parallel compiliation but nvlink is not bundled with Reactant's CUDA jll
2. There appears to be invalidation issues where a cached kernel is re-used instead of invalidated on change. This leads to issues that appear like buffer aliasing. Need to investigate exactly what is going on here, less confident about the actual diagnosis

Start by reproducing the persistent-kernel-cache path and checking whether Reactant's CUDA jll provides nvlink for parallel compilation. Then investigate cache invalidation when a kernel changes, using the reported buffer-aliasing-like failures as the symptom. Done means parallel compilation works and changed kernels are not incorrectly reused.

Written by the indexing model from the issue text.