Introduce RealtimeSanitizer (RADSan) real-time safety checking

[davidtrevelyan]

This PR does the following:

• Introduces a new GitHub Actions workflow that builds and tests RTNeural using RADSan
• Creates a new config.h header file, in which all RTNeural preprocessor definitions are contained
  • The existing definitions like RTNEURAL_NAMESPACE etc. have also been moved to config.h in an attempt to be clean, but I'm very happy to move them back to the RTNeural.h header if that's preferred!
  • Introduces the RADSan [[clang::realtime]] function attribute, configured behind a new RTNEURAL_REALTIME macro
• Enables the user (and CI!) to configure whether RADSan is being used with the cmake option -DRTNEURAL_ENABLE_RADSAN=ON
• Adds the configured attribute RTNEURAL_REALTIME to the real-time methods of Model and ModelT
• Fixes my earlier mistake that prevented failing test being output to the terminal in CI jobs

A couple of questions from me:

1. Is RTNEURAL_REALTIME a good enough name? Happy to change it to something else :)
2. Would it be a good idea to add RTNEURAL_REALTIME to the real-time methods of the Layer implementations?

[codecov-commenter]

Codecov Report

Attention: 14 lines in your changes are missing coverage. Please review.

Comparison is base (04cb333) 95.70% compared to head (c3103c9) 95.70%.

Files	Patch %	Lines
RTNeural/batchnorm/batchnorm.h	50.00%	2 Missing ⚠️
RTNeural/batchnorm/batchnorm2d.h	50.00%	2 Missing ⚠️
RTNeural/batchnorm/batchnorm2d_eigen.h	50.00%	2 Missing ⚠️
RTNeural/batchnorm/batchnorm2d_xsimd.h	50.00%	2 Missing ⚠️
RTNeural/batchnorm/batchnorm_eigen.h	50.00%	2 Missing ⚠️
RTNeural/batchnorm/batchnorm_xsimd.h	50.00%	2 Missing ⚠️
RTNeural/conv1d_stateless/conv1d_stateless_eigen.h	50.00%	1 Missing ⚠️
RTNeural/conv1d_stateless/conv1d_stateless_xsimd.h	50.00%	1 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #121   +/-   ##
=======================================
  Coverage   95.70%   95.70%           
=======================================
  Files          58       58           
  Lines        3891     3891           
=======================================
  Hits         3724     3724           
  Misses        167      167           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[jatinchowdhury18]

Looks good to me! Moving the macro handling to config.h is nice, and I think RTNEURAL_REALTIME is probably the right name for the attribute macro.

I guess we could add the realtime attribute to the layer-level forward methods. I don't think it would make much difference for the testing, but maybe it could be useful for folks who are using RTNeural layers without the "Model" API?

[davidtrevelyan]

Nice one, thanks @jatinchowdhury18 - agreed! I've just pushed up some additions of the attribute to the Layer methods that I think are intended to be real-time safe. Would you mind quickly double checking that I've understood them correctly, and haven't added the attribute to any methods that aren't intended to be real-time safe?

[jatinchowdhury18]

Sure thing! Your additions look good to me. I hadn't really thought much about whether the methods used to set the layer weights were realtime-safe, but it's cool that they are! I guess that could be important if there's a situation where someone wants to "modulate" their layer weights or something like that.

I'm ready to merge this PR if you are!

[davidtrevelyan]

Great! Yeah - super cool that the weights can be modulated - kind of like circuit bending for ML 🙃

Also, it could equally be really useful for controlling any highly interpretable differentiable elements that may live in the network, like a differentiable fader or something. Lots to play with, and I think RTNeural is really well placed in offering that option to creative developers!

I'm happy for this MR to be merged if you are! Thanks for the review 🙏