Bring in prefetch

[jeromekelleher]

After sgkit-dev/vcztools#408 we should have a decent perf uptick by reducing stalls in the hot path.

[jeromekelleher]

I did an analysis on this and it's a decent win for BGEN, and a very slight regression for plink.

format	chunk size	wall-time effect	per-advance stall
PLINK 4 KB	25 MB	+3 %	80 ms → 9 ms
PLINK 128 KB	25 MB	+4 %	unchanged (~200 ms)
PLINK qctool	25 MB	+9 to +29 %	already ~10 µs
BGEN 4 KB	300 MB	-14 %	500 ms → 25 ms
BGEN 128 KB	300 MB	-18 %	500 ms → 20 ms
BGEN 1 MB	300 MB	-22 %	600 ms → 500 ms
BGEN qctool	300 MB	-2 to -10 %	shaves ~20 %