Summary:
This is a complete reworking of the way that Haxl schedules I/O. The
main benefits are:
* Data fetches are no longer organised into rounds, but can be
arbitrarily overlapped with each other and with computation. The
scheduler supports an arbitrary queue of work items which it can
evaluate while data-fetching is taking place in the background. To
take advantage of this, data sources must implement a new form of
`PerformFetch`, namely `BackgroundFetch`. The old forms of
`PerformFetch` are still supported, but won't benefit from any
additional concurrency.
* It is now possible to specify on a per-data-source basis whether
fetching should be optimised for batching or for latency. A request
to a data source that doesn't benefit from batching can be submitted
immediately. This is done with the new `schedulerHint` method of
`DataSource`.
Reviewed By: niteria
Differential Revision: D4938005
fbshipit-source-id: 96f12ad05ee62d62474ee4cc1215f19d0a6fcdf3
