[Feature Request] Add option to cache model compilation for modular/max-openai-api

[ematejska]

Originally filed by remorses in #271. With the merge of max/mojo repo, reopening here.

What is your request?

I tried deploying modular/max-openai-api to fly.io, but it takes a lot of time to do the first compilation of the model, is it possible to cache the model compilation on disk?

What is your motivation for this change?

Add --model-compile-cache=/.root/model parameter

Any other details?

Fly.io is a serverless GPU deployment platform, the machine is stopped and started often, now model compilation is too slow to be able to deploy in this kind of infrastructure