[bestrq] stable training

[Mddct]

changes:

• stack fbank instead of stack features after subsampling （BEST-RQ paper do stack fbank ）
• replace compute_mask_indices_v2 with compute_mask_indices,(v2 from wav2vec2 which compute mask in cpu), The v2 version can consider padding, and the probability of short sequence overmask becomes smaller
• fbank should do cmvn ， stack fbanks should renorm using unlearnable LN avoiding codebooks collapse
• add fbank l2 regularization which can stable training
• other info about training
• reset encoder parameter using bert-style: trunk_norm std==0.02

[Mddct]

BTW: link error: ./test/test_tokenize.py:108:16: C419 Unnecessary list comprehension passed to all() prevents short-circuiting - rewrite as a generator. ' not in this pr

[xxchauncey]

hi, I guess xs here is just the log mel fbank feature from dataloader, so it seems that calculating features_pen is not meaningful in this case.