Faster interactive python workflow

[aaron-skydio]

Need to make this better, and then split out remaining items into additional TODOs

P1: The interactive workflow is too slow. The following 10 rotation chain takes 4.8 seconds to generate

2.9 seconds is rendering template, of which 2 seconds is calling black (lol), .8 seconds is calling string join inside of the template, and a small fraction is actually rendering

• We’re calling black.format_str 38 times. we should probably disable this for interactive use - flag in the language args of codegen?
  • I made an option to LanguageArgs to not autoformat. Perhaps controversial, but I’d propose to turn it off for the python Factor by default.
• Now str join is taking all the time of render template. What’s up with that?
• Why 38 times? We have 19 factors, so must be twice per factor.
  • What’s the second python file per factor? An __init__.py file from symforce/symforce/codegen/python_templates/function/__init__.py.jinja. Gross.
  • All prior and between factors should be the same expression. Can we only generate two?
• 2 seconds is in codegen_util.print_code. Seems like a lot of subs operations and list comprehensions, as well as some time in the sympy printer tree. Probably lots to optimize.

• A ton of time is being spent in subs in codegen_util.perform_cse, basically all initialization._get_subs_dict → _flatten_storage_type_subs → mostly python_util.scalar_like. We should work on reducing the calls to these or making them faster.

[aaron-skydio]

Fixed (for now) by the various performance improvements in v0.3.0