Draft: Introduce optional process pool feature to avoid GIL

[PhilippvK]

See #153 for more context

Summary of changes:

• Drop Run.session in favor of passing the session as argument when needed (Session class can not be serialized)
• Fix: broken detection of global config passed via cmdline (TODO: cherry-pick to develop)
• Add Run.has_target()helper
• Introduce MlonMcuContextMinimal to pass environment settings and cache to processes (Locks can not be serialized)

Limitations:

• runs_per_stage feature does not work together with process_pool
• No support for progress bars (--progress)
• Run.prefix property is missing session idx
• Run dir is created in /tmp instead of session dir

TODOs:

• [ ] Cleanups/Linting
• [ ] Documentation
• [ ] Testing
• [ ] Restore progress bar support
• [ ] Use RunInitializer and RunResult by default
• [ ] Refactor MLonMCUContextMinimal
• [ ] Expose session IDX and session dir to runs which out sharing session class
• [ ] Drop Run.__deepcopy__ as not needed anymore

[PhilippvK]

I gave this a try using a benchmark I was performing recently.

process_pool, per_stage=1: not supported
process_pool, per_stage=0: 53s
thread_pool,  per_stage=1: 15m22s 
thread_pool,  per_stage=0: 5m28s

I got a speed up of 5x over the ThreadPoolExecutor version on a 18C/36T machine which is great.

Visualization:

Scripts used to generate this plots can be found here: https://github.com/PhilippvK/mlonmcu-monitoring

[PhilippvK]

This PR has been superseded by the feature-process-pool-new Branch