Write specifications for all inferencing modules and align

[jfomhover]

There's a trade-off between aligning all inferencing modules, and providing flexibility to implement framework-specific arguments.

The inferencing pipeline should be able to rely on a standard set of parameters (ex: data type, batch size, num threads...) while providing flexibility for experiments requiring framework params (ex: using custom binaries).

Writing down a specification in our documentation, and decide which params should be the core "contract" of inferencing modules, then align on each of them, would put us in a better place before expanding to more frameworks.