Reload saved model

[christinab12]

Main change includes that now the model is saved and reloaded without errors and eval after train uses the latest model (before it was using the original model weights from init - train doesn't not update self.model, because it is async?).

Changes are:

• The saving of the model was changed from bentoml to bentoml.pytorch and all model classes now inherit from nn.module
• Both init and eval of the runner check if a model exists and load the latest version of it if so
• TEMPORARY WORKAROUND: mkldnn was disabled in cellpose as it was causing problems with saving and loading model (this also solves [BUG]Error: RuntimeError: itensor_from_mkldnn expects MKL-DNN tensor input when transitioning from evaluation to training mode MouseLand/cellpose#820)

[christinab12]

Thanks for the test fix @KorenMary!