[Mamba2] Post Merge Fixes - norm_before_gate and generation with inputs_embeds

[vasqu]

Since the implementation seems like a mirror to the one of the transformers repo, I thought I'd share some fixes post-merge that have been applied:

1. Generation when passing inputs_embeds has been broken, ref. Fix: Mamba2 generation mismatch between input_ids and inputs_embeds huggingface/transformers#32694
2. Incorrect usage of the norm_before_gate flag which causes misaligned training when using the completely fused kernel, ref. Fix: Mamba2 norm_before_gate usage huggingface/transformers#32686

Cool repo btw :)

[yzhangcs]

Thank you :D