Add more buffer pools

A follow-up to !1090 (merged), although with less impact due to these buffers being significantly smaller.

Still improves overall performance by around 0.5% or so on Windows by reducing the number of page faults:

Chimera (1080p):     91k ->  43k
HoliFestival (4K): 1343k -> 114k

No measurable impact on Linux.

Random trivia: The mutexes are only required due to potential reallocations. If sizes were guaranteed to not change a lockless implementation would be possible.