Mamba2 conversion script for original models (#32580)
* first attempt at allowing both conversions from codestral and from the original mamba ssm
* allow fp16, seems default for mamba2
* dtype fix
* simplify codestral check, dont overwrite pad/eos/bos when codestral
* change file -> directory
* use path join to be safe
* style
* apply code review
- add util mamba2 tokenizer (gptneox with left padding)
- add models dict
* fix copies
* add tokenizer to docs
* empty commit to check for weird err
* make conversion user dependent on model type, defaults for original paper models
* small comment nit
* remove norm_before_gate in conversion
* simplify model dict by using shared keys directly + remove unnecessary attributes
* fix tokenization: remove separate mamba2 tokenizer, add padding option as kwarg to gptneox one and reuse it for the conversion script
* simplify even further as we pass padding side via **kwargs already
