Main w2v2 pretraining

[artemru]

What does this PR do? Please describe: Merges MMS team's working fs2 branch into fs2 main. The PR contains

• dataloading implementations for w2v2 pretraining and ASR finetuning
• additional SpeechReadOptions fields
• LLM-ASR model and decoding implementation
• RNN-T ASR model and decoding implementation
• small changes to load_wav2vec2_asr_trainer() to support validating on multiple dev sets
• update wandb reporting to report metrics separately for all dev splits
• support specifying score_metric to determine which metric to use for dev- model selection in ASR

We use this with main branch of mms2 repo