sharedcache: move to fine-grained locking

Before this commit, s.mu was held when s.get & s.set were reading and
writing to s.file. This was done to simplify merging an initial implementation
with the necessary correctness properties. This commit moves to fine-grained
locking, to improve the performance of the cache. A table mapping each cache
block index to its "lock state" is introduced. During the reading and writing
to s.file, only the specific cache block being updated is locked. Both read and
write locks are supported, so that multiple reads of some cache block can be
done concurrently. Also, writing back to the cache is now done asynchronously
with respect to the call to ReadAt.

With this commit merged, I think we can run some production experiments
measuring the efficiency of the cache.