Allow FSDP prefetch to be configurable

[RobotSail]

To optimize training times, FSDP allows you to overlap computation by either prefetching parameters before the current set of parameters' gradient computation. While fast, this adds to the total memory used by FSDP. There are three total prefetch policies provided by FSDP, which are documented here

Since this setting affects the total memory used by FSDP, we should allow this to be a configurable setting within the FSDPOptions object.