Make MPI initialization optional and use OMPI env vars for world discovery

[jfomhover]

The initialization of MPI in python might interact with the initialization done by the framework themselves (ex: lighgbm training). We might still need it in some scenarios (sharing distributed metrics like in #185), so we're making this optional for now.

When mpi_init_node is left to None, we use OMPI env variables to discover how many nodes there are, and create the right configuration for the distributed script to use.

We're also proposing to upgrade to openmpi4.

NOTE: this is part of an effort to try this distributed training on production scenarios where we're hitting internal mpi exceptions (MPI_ERR_TRUNCATE).

[github-actions]

Unit Test Results for Build

  1 files  ±0    1 suites  ±0   33s ⏱️ +3s
91 tests +1  91 ✔️ +1  0 💤 ±0  0 ❌ ±0 

Results for commit 22285b2. ± Comparison against base commit b1bea5e.

This pull request removes 1 and adds 2 tests. Note that renamed tests count towards both.

tests.common.test_distributed ‑ test_mpi_handler

tests.common.test_distributed ‑ test_mpi_handler_mpi_init
tests.common.test_distributed ‑ test_mpi_handler_no_mpi_init

[github-actions]

Package	Line Rate	Branch Rate	Complexity
common	85%	0%	0
scripts	100%	0%	0
scripts.data_processing.generate_data	93%	0%	0
scripts.data_processing.lightgbm_data2bin	95%	0%	0
scripts.data_processing.partition_data	92%	0%	0
scripts.inferencing.custom_win_cli	94%	0%	0
scripts.inferencing.lightgbm_c_api	75%	0%	0
scripts.inferencing.lightgbm_python	95%	0%	0
scripts.inferencing.treelite_python	94%	0%	0
scripts.model_transformation.treelite_compile	92%	0%	0
scripts.sample	93%	0%	0
scripts.training.lightgbm_python	80%	0%	0
Summary	86% (1227 / 1426)	0% (0 / 0)	0