[Snapshot Performance] Batch accesses

[ripatel-fd]

The size of Firedancer's account index easily exceeds L3 cache size.
A simple loop that inserts a billion accounts into a funk index would thus be stalled frequently on DRAM fetches.

(Easiest to diagnose on an Intel CPU but I found some performance counters that we can use for AMD too)

Techniques like scatter-gather or prefetch instructions should allow for noticeable performance improvements when the list of addresses to query is known in advance.