[Reporting Tool] Modular report classes for saving non-Thunder-specific repro scripts

[kiya00]

For the next PR, I think it'd be useful to create an FXReport object. What I'm thinking is:

• Create a new function called fx_report (not Thunder-specific)
• Return a FXReport object from fx_report
• The FXReport object has a sequence of FXGraphReport objects, one per fx graph in the original program
• Each FXGraphReport has a method to write a reproducer script
  • The method should be able to easily select whether the reproducer uses PyTorch eager, torch.compile, or thunderfx (with some supported options)
  • For total control, the method should allow writing an arbitrary executor string and allow for extending what's imported with arbitrary import strings, in case someone wants to use a very special version of thunderfx or torch.compile or PyTorch eager
• Each FXGraphReport should allow its FXGraph's inputs to be queried (they can just be FakeTensors for now)

So someone could do something like:

report: FXReport = fx_report(foo)

fxgr: FXGraphReport
for fxgr in report.graph_reports:
  print(fxgr.inputs)
  fxgr.write_eager_reproducer(filename0)
  fxgr.write_torchcompile_reproducer(filename1)
  fxgr.write_thunder_reproducer(filename2)
  fxgr.write_reproducer(filename3, executor_string="my_crazy_executor", import_strings=("import my_crazy_module", "from my_crazy_module import my_crazy_executor"))

And that would generate four reproducers for each of the FX graphs in the program, each one using the different requested executor.

Writing the reproducers directly like this may not be super interesting in isolation, but once that's working we can add more of the data that thunderfx_report makes available, and that will become really interesting really fast. It would also be great if the existing reporting functionality, like thunderfx_report could be updated to use FXGraphReport and FXReport objects internally for consistency. Using modular Python objects for the report data and analysis should make it easier for us to develop and for others to extend.

What are your thoughts, @kiya00?

Originally posted by @mruberry in #1636 (comment)