Memory reduction fixes for MCMC sampler

[andrewdipper]

Here are some proposed changes for reducing GPU memory footprint for MCMC sampling. I was hitting OOM earlier than expected so tracked some of it down:

• Collection was updated within jit blocks without donate_argnums (at least when using a progressbar). This added a copy for each update. Using donate_argnums makes sampling time with a progress bar much closer to without as well.
• Unraveling the flattened collection array at the end requires a copy. I modified collection to keep the pytree structure during sampling to avoid the replication.
• Waiting to materialize _states_flat. Since jax doesn't use views for reshaping this duplicates all the samples. _states and _states_flat might even have the same memory layout. I made a rough change to not materialize _states_flat until requested but I'm not sure that's an ideal solution
• The cached_by change was to allow saving multiple functions within fori_collect. Otherwise they'd collide and use the wrong function.

Below are some rough numbers for peak memory usage / runtime with and without the changes for two models (split by //) just to give an initial view.

baseline / no progress bar: 3592MB / 63sec
baseline / with progress bar: 5126MB/ 109sec // 14412MB / 515sec new / no progress bar: 2052MB / 63sec new / with progress bar: 2052MB / 70sec // 5182MB / 320sec

Let me know if you think any of the changes would be useful / any modifications are needed

[andrewdipper]

Is there any way to get the output / insight into the failing test? FAILED test/test_examples.py::test_cpu[stochastic_volatility.py --num-samples 100 --num-warmup 100] - subprocess.CalledProcessError: Command '['/opt/hostedtoolcache/Python/3.9.19/x64/bin/python', '/home/runner/work/numpyro/numpyro/examples/stochastic_volatility.py', '--num-samples', '100', '--num-warmup', '100']' returned non-zero exit status 1.

For the latest update I've run the tests both locally and on a kaggle T4x2 session and they passed. The prolda test sometimes has issues with getting the dataset but I've had no problems with the stochastic_volatility test.

[andrewdipper]

I accidentally updated this with a merge instead of rebasing - let me know if that's an issue. Anyhow it's back up to date

[andrewdipper]

For sure, makes sense

[fehiepsi]

Happy to merge! 💯