Add DPTNet and WarmupStepLR scheduler

[Emrys365]

This PR adds the dual-path transformer network (DPTNet) for time-domain speech separation, and the corresponding LR scheduler (I call it WarmupStepLR here) proposed in its paper.

[codecov]

Codecov Report

Merging #4449 (9be1554) into master (986dadb) will increase coverage by 0.00%.
The diff coverage is 86.98%.

@@           Coverage Diff            @@
##           master    #4449    +/-   ##
========================================
  Coverage   82.77%   82.77%            
========================================
  Files         474      477     +3     
  Lines       40755    40906   +151     
========================================
+ Hits        33733    33860   +127     
- Misses       7022     7046    +24     

Flag	Coverage Δ
test_integration_espnet1	66.58% <ø> (ø)
test_integration_espnet2	49.17% <26.71%> (-0.14%)	⬇️
test_python	69.78% <86.98%> (+0.05%)	⬆️
test_utils	23.45% <ø> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
espnet2/schedulers/warmup_step_lr.py	34.61% <34.61%> (ø)
espnet2/enh/separator/dptnet_separator.py	96.72% <96.72%> (ø)
espnet2/enh/layers/dptnet.py	100.00% <100.00%> (ø)
espnet2/tasks/abs_task.py	75.44% <100.00%> (+0.03%)	⬆️
espnet2/tasks/enh.py	99.23% <100.00%> (+<0.01%)	⬆️
espnet2/main_funcs/calculate_all_attentions.py	95.52% <0.00%> (-4.48%)	⬇️
espnet2/enh/separator/tcn_separator.py	96.77% <0.00%> (-3.23%)	⬇️
espnet2/enh/separator/dc_crn_separator.py	95.00% <0.00%> (-2.50%)	⬇️

📣 Codecov can now indicate which changes are the most critical in Pull Requests. Learn more

[Emrys365]

Hi @popcornell Could you help review this PR?

[popcornell]

It looks good to me already.

[Emrys365]

The trained model has been uploaded to HuggingFace: https://huggingface.co/espnet/Wangyou_Zhang_wsj0_2mix_enh_train_enh_dptnet_raw

[Emrys365]

I think this PR is ready to merge.

[popcornell]

LGTM !

[Emrys365]

Thanks! @popcornell
I will merge it.