Cache ctx.saved_tensors in Mamba3 backward to fix FSDP activation checkpointing

[anasiri]

_Mamba3Function.backward() accesses ctx.saved_tensors twice. torch.utils.checkpoint registers unpack hooks that only allow a single access, so the second call raises CheckpointError when FSDP activation checkpointing is enabled:

torch.utils.checkpoint.CheckpointError: torch.utils.checkpoint: Unpack is being triggered for a tensor that was already unpacked once. If you are calling ctx.saved_tensors in backward, make sure to do so only once. Otherwise please open an issue with details on your use case.

This PR caches ctx.saved_tensors in a local variable so both uses read from the same reference.