Add CPU offload strategy for model (auto)unload

[brycehenson]

adds a unload strategy which moves the model weights to cpu ram. This substantially speeds up the time to get the model ready for a request.

      - MODEL_UNLOAD_STRATEGY=move_to_cpu

default is "destroy" which reloads from disk

Benchmark

RTX 2070 SUPER with PCIe Gen3 x16 link

Timing

strategy	unload avg	unload stddev	load avg	load stddev
destroy	0.027s	0.015s	0.979s	0.189s
move_to_cpu	0.299s	0.022s	0.078s	0.007s

Memory

Memory

strategy	VRAM before unload	VRAM after unload	VRAM delta	service RSS delta	system RAM delta
destroy	1122.800 MB	630.800 MB	-492.000 MB	-20.663 MB	-22.962 MB
move_to_cpu	915.600 MB	715.600 MB	-200.000 MB	+236.111 MB	+180.091 MB

[RBEmerson970]

adds a cpu cache of the model to reduce load speed

Reduce load speed, or reduce load time? I don't understand why reducing load speed would be a good thing.