feat: Allow for more flexible shapes

[sebffischer]

Previously, only the batch dimension was allowed to be NA.
However, this assumption is too restrictive, as it does not allow for transformer-based architectures where the sequence dimension is unknown.
This PR changes this and NAs can now be at every position of the shape.
However, many nn() operators expect only the batch dimension to be unknown.
Therefore, the only_batch_unknown argument was added to PipeOpTorch that is TRUE by default and needs to be overwritten.
E.g. PipeOpTorchLinear can handle NAs, as long as they are now in the last dimension.
Other operators that can handle it are, e.g., activation functions or in the future the multihead-attention module.

This PR also improved the shape inference and fixed some other small bugs.

TODOs:

• the infer_shapes() method must be adjusted. Now, we replace all NAs dimensions with an arbitrary dimension.
  we should also do this twice to check whether the results are compatible
• check all occurences of assert_shape() and verify that we no longer make the assumption that only the batch dimension is NA
• check the PipeOpTorch object implementations that they err gracefully --> PipeOpTorch now need to indicate whether they can handle NAs that are not in the batch dimension.
• Update the documentation on the shape

[sebffischer]

Suggested change

	* feat: Shapes can now have multiple `NA`s and not only the batch dimension can be missing. However, most `nn()` operators still expect only one missing values and will throw an error if multiple dimensions are missing.
	* feat: Shapes can now have multiple `NA`s and not only the batch dimension can be missing. However, most `nn()` operators still expect only one missing values and will throw an error if multiple dimensions are missing.
	* feat: better inferring of output shapes via function `infer_shapes()`.