Pipelined/concurrent read for startup

[dgrnbrg]

This improves application startup by allowing for parallel startup. Currently, we interleave reading hints/entries from the files on application startup with processing them into keydir entries and loading them into the keydir. These could be done on 2 threads: one for IO, and one for parsing. Also, the files are read sequentially, when really they could be read concurrently to further improve startup time.

If they are read concurrently, we'd need to choose the latest value for the keydir based on the file's timestamp, since the unix timestamps stored in the file are of insufficient resolution.