[fuzz] a patchable bug filter

[ganler]

• defining "a filtered bug report":
  • a bug report will not be logged or dumped;
  • nevertheless, it is still a bug -- it's just the user wants to ignore it for some reason.
• native support: nan, inf, dup (duplicate filter with string hash)
  • + fuzz args: filter.type="[nan,inf]"
• bring your own filter:
  • implement a callable to (report: BugReport) -> bool where True means filtering the bug.
  • put the file to path/to/patch.py
  • decorate it with nnsmith.filter.filter("filter_name") (note nan, inf, dup are taken)
  • load it at runtime: + fuzz args filter.type="[nan,inf,filter_name]" filter.patch=/path/to/patch.py
• tested patching with a dummy patch (tests/mock/filter_patch.py) in the end-to-end fuzzing script in the ci.

[ganler]

Did some longer fuzzing (8hr) for torch-onnx side. I decide to enclose 0.0.1 if tensorflow side can also stably run for a relatively longer term (say at least 4hr), given the recent improvements (16 PRs/squashed commits).

[ganler]

It seems for xla and tflite, some crash bugs can happen within 2 hours. That said before 0.0.1 release, I will implement crash safety asap.