Use thread local storage for frontend compile cache

[zcbenz]

Refs #2086, #3078, #3216.

The frontend compile cache uses function ID as key so it is very unlikely functions in different threads would shared cache, and making it thread local should be enough to ensure thread safety.

For backend compile cache, i.e. Compiled::eval_cpu/eval_gpu, it is reasonable to share cache between threads and they are already guarded with locks (except for the CUDA backend which I will fix in followup PRs).

[angeloskath]

Looks great!

I wonder if people rely on compile being ran once. They shouldn't but this could be a fun bug to debug in the future.

@mx.compile
def fun(x):
    call_my_run_once_and_only_once_side_effect_op()
    return x**2 + 2 * x + 1

# thread 1
fun(mx.ones((10,))

# thread 2
fun(mx.ones((10,))) # oops

[zcbenz]

That would already be a problem since changing the default stream would invalidate the compile cache. In a multi-thread environment we expect users to use different streams as default streams so hopefully it would not be a problem.