Add PieceCache::read_pieces() method for more efficient piece cache reads in batches

[nazar-pc]

This builds on #3135 and #3141 to implement more efficient piece reading.

For local cache we now have a pool of open files. DirectIoFile has mutex internally due to need to manage its read/write buffer for aligned I/O and without pool concurrent reads are not actually happening despite &self of the FileExt trait it implements. Also it is from our experience with farmer faster to read from multiple handles in general, especially on Windows (though due to memory usage issues Windows is still doing blocking non-concurrent reads).

For cluster cache we now have stream request for faster retrieval of pieces, which was much easier to implement after refactoring done in #3141

I also clarified guarantees about returned values in a few places.

The biggest remaining topic is implementing PieceGetter::get_pieces() for ClusterPieceGetter and remove dummy implementation from PieceGetter.

Not particularly happy with number of hashmaps required in piece cache handling, but I didn't want to force all results being returned in-order either and without guaranteed order we need an efficient way to track pieces, hence hashmaps 😞

Code contributor checklist:

• I have read, understood and followed contributing guide

[teor2345]

Looks good, I just had a question about a trait method invariant.