Added dai_type, updated encoding flags, and removed reverse_input_channels flag

[ptoupas]

Main changes:

• Updated the encoding.from and encoding.to flags to replace the reverse_input_channels.
• Removed reverse_input_channels.
• Added dai_type (its still not 100% compatible with all DAI data types).
• Fixed a bug with mean and scale values on RVC2 conversion with multiple inputs.
• Removed the extra nodes added to the onnx model when mean=[0,0,0] and scale=[1,1,1]
• Added extra unit tests for the encoding

[kozlov721]

In general LGTM, just left a comment regarding removing the reverse_input_channels. I'd propose keeping it as a property of InputConfig, somehting like:

@property
def reverse_input_channels(self) -> bool:
    return self.encoding.to != self.encoding.from_

That way we don't need to change the rest of the code where inp.reverse_input_channels is used. And it also looks much more readable than doing inp.encoding.to != inp.encoding.from_ everywhere.

[klemen1999]

Generally looks good. I left some comments but mainly what we are missing is rigorous testing. By that I mean:

• test of the output NNArchive for different input configs, intermediate NNArchives and/or CLI params
• tests of intermediate onnx (for RVC4) to check which blocks are added and if values inside them are correct (mean and scale properly reversed if need be)
• test that flags for RVC2/3 and HAILO conversion are properly set up depending on the input config
• any other tests that will help catch any current or future bugs

[klemen1999]

LGTM, only thing left to update is the requirements.txt to luxonis-ml>=0.5.0
CC: @kozlov721 to re-review

[kozlov721]

LGTM