- Makes the safetensors module private.
  - Doesn't get exported on the preamble, avoiding a naming clash with the safetensors external crate.
- Change how and when the period is inserted.
  - This should make it closer to how the fields are accessed in the code.

- Add the try_normalize_rms related functions.
- Add the `LayerRMSNorm1D` module.

- Add `TrySplitShapeAlong` and `TrySplitTensorAlong`.
- Minor linting and docs fix.

TODO
- Check if the tape should be returned. If not, it can be removed from the interface.
- Add cuda kernel.
- Consider a different interface, where it could get split in more than two tensors - possibly stated on a vec.
  In this way it could get closer to the pytorch interface (chunks).

- Also added `from_fn` for Arrays.

Note: the interface currently requires two passes for construction, one for creating a list of tensors with NoneTape and another for putting tapes into those tensors.

Remove ftz

This alternative method:
- Requires load/read to decide whether it should skip missing tensors;
- Requires load/read/save/write to decide how should keys be mapped.

- Add stateless forward impl.
  - Efficient for training (but training is not yet implemented).
  - Input requires the entire sequence, and requires no state cache.
  - Generates one output for each input sequence.
- Add stateful forward impl.
  - Efficient for inference.
  - Input requires the last single sequence point, and requires the last state cache.
  - Generates a single output referring to the last input.