Better concurrency when lowering

[Xanewok]

Originally created by @nrc (rust-dev-tools/rls-analysis#157):

We do a lot of locking (every time we consult the project analysis) when lowering. We should be able to do better, at the least by using RW locks, but perhaps by being smart and doing some lock-free magic or some cloning

[Xanewok]

Originally replied by @nrc (rust-dev-tools/rls-analysis#157 (comment)):

We could also use more fine-grained locking, at present we have a single lock over the whole analysis which is taken and dropped repeatedly. It would be more efficient to have finer grained locks and hold them for longer.

Given we don't actually make use of parallelism, we could just hold the lock for the whole of lowering (this is a bit tricky because we use the user-facing API which includes locking.