Add `cuda.core.utils.PersistentProgramCache`

[leofang]

Two kinds of on-disk caches:

• one for sqlite3 based (suitable for single processes)
• one for file-stream based (suitable for concurrent processes, thanks to atomic I/O)

It is unclear to me (yet) if they can be easily combined under the same class PersistentProgramCache. We'll need some research.