rate limit

[gnurizen]

Ignore first commit, that's PR #6

• Implement simple rate limiting scheme to not overwhelm the server

[gnurizen]

Gotta say I don't love this:

1. There's plenty of bandwidth via ebpf perf events so we could easily push this into the agent instead of the shim.
2. With some workloads we can miss important events (like rare but higher impact graph launches in a sea of individual kernel launches)
3. With bursty activity that isn't constant we can miss a lot of stuff (inference workloads)
4. We can still bust the server if we have a graph launch that has many kernels as we'll generate a sample for each kernel under the 1 graph launch (in practice I usually only see on the order of a dozen or so kernels per graph launch but I suspect it can get into the hundreds).

That said we need some form of rate limiting in the shim because some workloads will generate so much unwind activity that we'll exceed desired CPU limits from that alone.

So I think I'm gonna keep this but change it so that the rate limit isn't applied to graph launches.

I'm also gonna add a env switch to disable it for testing/diagnostic purposes.