[christophe0606]

Goal of the PR:

• We may have several methods in the same .pte to support more complex data pipelines
• We may need metadata in the .pte to have a code that is more generic
• We may need different .pte for different data pipelines
• Ethos-U quantization can use different scheme (get_symmetric_quantization_config or get_symmetric_a16w8_quantization_config)
• We may work with networks that have already been quantized

To support those use cases, we need a more complex contract between create_ai_layer.py and model.py and probably a rework of the model node in mlops.

The new python description.py provides datatype to describe more complex use-cases.
model.py can be used as before and nothing will change. But, by using the new datatype it can describes different needs to create_ai_layer.py.

It introduces new classes : Model, Metadata, Pte

The Model class describes:

• Different name for the model
• If the model is already quantized or not and the function to use to provide calibration samples
• If the model needs a custom quantization (not default one from create_ai_layer) and the function to use to do this custom quantization
• Information like input-shape and calibration-samples no more read from the mlops

A Metadata class to provide metadata in the pte

A Pte class to package several models + metadata in the pte

A list of Pte object to generate several Pte.

The get_model function can thus generate:

1. A nn.Module as before
2. A Model
3. A list of Model
4. A Pte containing one or several Model
5. A list of Pte

The generated ai_layer will contain all operators needed by all networks of all pte.
There will be a different group per pte.

Not covered:

1. Multiple NPU support
2. Custom vela options per .pte
3. Future vela features like VelaExternalBlockPlacements
4. More customization for memory management

Reference issue : #41

[christophe0606]

I'll probably move some information given through the dataclasses to the mlops description in csolution ... so this PR is a work in progress