Zero_infinity option for ctc loss

[kamo-naoyuki]

When I implemented ignore_nan_grad for ctc modules of espnet2, pytorch-ctc didn't have the zero_infinity option.

ignore_nan_grad is not so efficient and looks complicated impelemetation, so I changed it to use the zero_infinity option. (ESPnet1 already used it.)

Note that this CTC class has both ignore_nan_grad and zero_infinity options and ignore_nan_grad is an alias of zero_infinity. I left ignore_nan_grad for backward compatiblity.

[codecov]

Codecov Report

Merging #4415 (cdea46d) into master (45e8cb9) will increase coverage by 0.01%.
The diff coverage is 83.33%.

@@            Coverage Diff             @@
##           master    #4415      +/-   ##
==========================================
+ Coverage   82.55%   82.56%   +0.01%     
==========================================
  Files         469      469              
  Lines       40228    40213      -15     
==========================================
- Hits        33211    33203       -8     
+ Misses       7017     7010       -7     

Flag	Coverage Δ
test_integration_espnet1	66.58% <ø> (ø)
test_integration_espnet2	49.55% <50.00%> (+<0.01%)	⬆️
test_python	69.36% <83.33%> (+<0.01%)	⬆️
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/asr/ctc.py	95.08% <83.33%> (+8.23%)	⬆️

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

[sw005320]

LGTM.
Thanks, @kamo-naoyuki