Improve mutex/lock handling

[destefy]

Currently, the compiler and inference thread of the remote-compiler client backend share a lock to ensure that the compiler can safely delete the uncompiled model once compiled is done.

However, this could lead to starvation if a thread holds onto a lock for too long.

We should probably use semaphores instead or whatever is the standard implementation for the reader-writer concurrency problem

┆Issue is synchronized with this Notion page by Unito