device-map test

[rsxdalv]

#32 for @nullnuller

This is an initial draft, further testing by different users is required. The code can also be written more elegantly, but this should be a robust way to see if it works and how good is the performance of this approach.

The code aims to 1. bring input tensors to self.ace_step_transformer.device and then bring the output tensors to self.device for each { ace_step_transformer, music_dcae, text_encoder_model }
It modifies cpu offload to follow the rules of self.device_map

Example device_map:

device_map = {
    'ace_step_transformer': "cuda:0",
    'text_encoder_model': "cuda:1",
    'music_dcae': "cuda:1",
}