llama : allow other bufts when overriding to CPU, add --no-repack option

[slaren]

• When using --override-tensor to override to the CPU, other buffer types will be considered as well. In practice, what this means is that the host buffer types will be used, which may improve performance when prompt processing is offloaded (Note that mmap needs to be disabled to use host buffers).
• Adds --no-repack (-nr) option to disable weight repacking.

llama-bench -m Qwen3-30B-A3B-Q4_0.gguf -ot exps=CPU -n 0 -p 32,64,128,256,512,1024 -ub 1024 -mmp 0:

Model	Test	t/s master	t/s sl/ot-repacking	Speedup
qwen3moe 30B.A3B Q4_0	pp32	15.03	22.62	1.50
qwen3moe 30B.A3B Q4_0	pp64	28.87	45.04	1.56
qwen3moe 30B.A3B Q4_0	pp128	61.06	89.35	1.46
qwen3moe 30B.A3B Q4_0	pp256	121.44	173.97	1.43
qwen3moe 30B.A3B Q4_0	pp512	227.41	309.59	1.36
qwen3moe 30B.A3B Q4_0	pp1024	421.50	594.32	1.41