feat: Allow sampling in the backgound

[aseyboldt]

Allow sampling in the background:

import pymc as pm
import nutpie

with pm.Model() as model:
    pm.Normal("x", shape=10_000)

compiled = nutpie.compile_pymc_model(model)

sampler = nutpie.sample(compiled, blocking=False)

try:
    sampler.wait(timeout=2)
except TimeoutError:
    pass

#sampler.abort()
trace = sampler.finalize()

cc @ricardoV94

Also fixes #90

[ricardoV94]

Neat

[ricardoV94]

Is there a way to probe if the sampling is finished without interrupting or blocking? Not clear from skimming the code

[aseyboldt]

Is there a way to probe if the sampling is finished without interrupting or blocking? Not clear from skimming the code

Now there is a is_finished property...

[aseyboldt]

And now also sampler.pause() and sampler.resume() and sampler.wait(timeout=seconds).

A form of sampler.inspect() that returns the partial trace that's been produced so far would stilll be nice, but I'll leave that for later.

[aseyboldt]

@ricardoV94 If you have a minute, this should be ready for review with the new changes again.

[ricardoV94]

Wanna test some of those methods in the new test?

---

Would be nice to setup some docs to show this functionality, but need not be in this PR.
The only docs that exist right now are the README :D
Perhaps a pymc-examples is also a good place.

[ricardoV94]

This is awesome. The inspect thing will also be great in the future!

[AlexAndorra]

This is super cool @aseyboldt !!
I just started playing around with this, but I think the finalize doesn't exist, so you can't do trace = sampler.finalize() to store the trace once the sampler is done.

What I ended up doing is:

compiled_model = nutpie.compile_pymc_model(model)
sampler = nutpie.sample(compiled_model, blocking=False, save_warmup=False, store_unconstrained=False)
idata.extend(sampler.abort())

That works, but feels weird from a user perspective.

Is this a bug? I'm happy to open an issue if that's helpful