Add iFaSNet for enhancement/separation tasks.

[LiChenda]

This is an implemntation of iFaSNet.
The code is based on Yi Luo's implementation.
I also add a corresponding configuration file for CHiME4 recipe. A pretrained model on CHiME4 will be uploaded to hugginface in near future.

[Emrys365]

Could you also add the original FasNet model?
They are based on the same codebase, and it would be convenient to have both models in ESPnet.

[codecov]

Codecov Report

Merging #4130 (d3d65cf) into master (a04a98c) will increase coverage by 0.16%.
The diff coverage is 93.20%.

@@            Coverage Diff             @@
##           master    #4130      +/-   ##
==========================================
+ Coverage   80.43%   80.60%   +0.16%     
==========================================
  Files         442      450       +8     
  Lines       38557    39372     +815     
==========================================
+ Hits        31015    31735     +720     
- Misses       7542     7637      +95     

Flag	Coverage Δ
test_integration_espnet1	67.13% <ø> (ø)
test_integration_espnet2	50.29% <14.88%> (-0.85%)	⬇️
test_python	66.98% <93.20%> (+0.46%)	⬆️
test_utils	24.45% <ø> (ø)

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

Impacted Files	Coverage Δ
espnet2/enh/layers/ifasnet.py	90.27% <90.27%> (ø)
espnet2/enh/layers/fasnet.py	92.85% <92.85%> (ø)
espnet2/enh/separator/fasnet_separator.py	96.29% <96.29%> (ø)
espnet2/enh/layers/dprnn.py	98.54% <96.36%> (-1.46%)	⬇️
espnet2/tasks/enh.py	99.08% <100.00%> (+<0.01%)	⬆️
espnet2/asr/decoder/transformer_decoder.py	95.08% <0.00%> (-1.56%)	⬇️
espnet2/tasks/asr.py	91.20% <0.00%> (-0.75%)	⬇️
espnet2/bin/mt_inference.py	0.00% <0.00%> (ø)
espnet2/asr/decoder/mlm_decoder.py	100.00% <0.00%> (ø)
espnet2/asr/encoder/longformer_encoder.py	80.62% <0.00%> (ø)
... and 7 more

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

[sw005320]

LGTM.
@Emrys365, if you're OK, you can marge it.

@LiChenda, it would be great if you prepare a pre-trained model and result later.

[sw005320]

LGTM.
@Emrys365, if you're OK, you can marge it.

@LiChenda, it would be great if you prepare a pre-trained model and result later.

[LiChenda]

Thanks for the comments! As @Emrys365 suggested, I also add the origin fasnet. The pretrained model will be uploaded soon.

[Emrys365]

Thanks a lot, @LiChenda