Fix regression of model loading performance when using mlock.

[spencersutton]

Performance of subsequent runs when using mlock suffered a large regression from 2347463.

It appears that changing the mmap flag from MAP_SHARED to MAP_PRIVATE causes the model to be loaded into memory fresh each time. This PR changes llama_mmap to take a new parameter called has_lora which if true calls mmap with MAP_PRIVATE and PROT_READ | PROT_WRITE otherwise uses MAP_SHARED and PROT_READ.

Caveats:

• I've only tested this on mac OS
• I haven't tested extensively actually using a LoRa
• Maybe use_mmap/use_mlock/has_lora should be combined into an enum.
• There is probably a better name for the parameter than has_lora.

[howard0su]

I will revert the change. I didn't test the mlock before I commit.

[ggerganov]

Do we still want this PR after the revert?

[howard0su]

No.

[mofosyne]

closing. Observed #2206 as the fix that renders this PR obsolete.