Adds training script

[ilyakava]

Hi and thank you for sharing the DPTNet code. I've added a training script that you will recognize from: https://github.com/ShiZiqiang/dual-path-RNNs-DPRNNs-based-speech-separation

The solver.py is basically copy-pasted from that repo, as well as their train.py -> train_and_eval.py here.

I kept all your parameters the same in the default model. I also added a logger that prints and saves to file, and updated the readme with example runs. There are also a few more cmd line arguments added for resuming training from saved checkpoints.

I'm training this now, batch size 1 with wsj0-2mix on 1x3090 takes about 12 epochs per day. Is this in line with your training procedure?

I have found that larger batch size leads to worse performance, this means that it must be trained on 1 device. Is this also in line with your experience?

[ilyakava]

I was able to train this for 100 epochs once and got a val loss of -20.156 on wsj0-2mix

[ilyakava]

I trained this a second time for 100 epochs and got training loss down to -21.109 and eval loss down to -20.229 on wsj0-2mix