Prevent GBM memory leaks by forcing actor cleanup after each epoch

[arnavgarg1]

Ray Actors get re-initialized after each epoch/boosting rounds, but the actors leave a lot of garbage hanging around in Ray's object store. This results in a few problems:

1. Increased in-memory data usage epoch after epoch since artifacts/objects aren't cleaned up, potentially leading to OOMs
2. Increased disk-spillage from object store to disk once the 20% threshold has passed, increasing time per epoch by an order of magnitude after each epoch

Together, these result in models either OOMing, or taking forever to train.

This PR forces cleanup after each epoch to prevent memory leaks by wrapping the LightGBM train step when using Ray with a ray.remote(max_calls=1) to kill the Ray Actors and GC at the end of every epoch. This involves two changes:

1. The creation of a modified train_loop function from the base class that enables making ray.remote calls.
2. Moving the train_step function outside of the LightGBMRayTrainer class and making it its own ray remote function that is called by the train_loop function inside the LightGBMRayTrainer class

[github-actions]

Unit Test Results

    4 files  ±    0      4 suites  ±0   1h 52m 33s ⏱️ + 1h 11m 17s
185 tests +152  156 ✔️ +126  29 💤 +26  0 ❌ ±0 
197 runs  +131  166 ✔️ +106  31 💤 +25  0 ❌ ±0 

Results for commit e550a58. ± Comparison against base commit cf5c604.

[justinxzhao]

Nice fix!