multi workers for valid dataloader

argparse load conf from file

format code

rm useless swig dir

refactor raw ctc decoder into ctcdecoder, new join ctc/att decoder

using simple test for multi decode type, and using gpu

update the benchmark

update timit result

update timit result

resume not increase epoch and step, before_train do it

fix ctc loss not convergence

Add bentchmark

Update the download scripts