HubAI conversion + deprecate blobconverter + ConvertOnTrainEnd callback that exports and archives

[dtronmans]

Purpose

• HubAI conversion: this is a new exporter, it takes as argument the platform and all other conversion parameters (eg: number_of_shaves) can be passed in the params
• Deprecate blobconverter: heavily discouraged through warnings
• Add ConvertOnTrainEnd callback: this performs export + archive
• Discourage the use of both ExportOnTrainEnd and ArchiveOnTrainEnd simultaneously and encourage the new ConvertOnTrainEnd callback

Specification

• The convert method runs both export and archive unconditionally: it is no longer possible to separately run an export step and an archive step. This shouldn't be a problem, I realized that we always "forced" users to have both ArchiveOnTrainEnd and ExportOnTrainEnd callbacks active thanks to smart_auto_populate(), so isolated use of either ArchiveOnTrainEnd or ExportOnTrainEnd wasn't really happening in the first place.
• The export method no longer runs the blobconverter
• ConvertOnTrainEnd without an exporter defined: generates the ONNX and the NNArchive ONNX only
• Convert as a CLI command: as an example you can do luxonis_train convert --config fruit_detection_config.yaml --weights output/51-salmon-dingo/best_val_metric/bestfruit.ckpt exporter.hubai.platform rvc4
• Small mistake that I made during the PR for unique ONNX identifiers, back then I changed it from a param in the method to a config parameter but forgot to remove the docstring parameter (this is fixed in this PR)

Dependencies & Potential Impact

Deployment Plan

Testing & Validation

As a callback at the end of training:

exporter:
  data_type: fp16
  hubai:
    active: true
    platform: rvc2
    params:
      number_of_shaves: 8

CLI:

luxonis_train convert --config fruit_detection_config.yaml --weights output/51-salmon-dingo/best_val_metric/bestfruit.ckpt exporter.hubai.platform rvc4

[klemen1999]

Left some initial comments, didn't actually test the code yet though

[klemen1999]

We'll wait a bit on this because this/next week a change in HubAI will come where INT8INT16 precision will get introduced. After that we'll properly adjust here as well if needed (TBD)