feat: lazy tensor datatype

[sebffischer]

We need this from mlr3pipelines: mlr-org/mlr3pipelines#737

[sebffischer]

How to solve the issue with the cloning for the preprocessing eventually:

We can create a special PipeOp / Graph class where there is only a $forward() method and no state / parameter values.
I.e. the use-case of PipeOpModule.

We can then maybe use this for the internal graph to do away with all the deep clones that we don't need in 99% of the cases.

[sebffischer]

Maybe: when augment is TRUE, input shape must be equal to output shape (are there cases where this is a restriction?)

[sebffischer]

Things to do:

• autotest should check that shapes applied to batch is coherent, i.e. giving NA and a specific value is identical but for the first dim
• Check that the predict shapes are added correctly and that it also throws an error (only use public API for this test)
• replace the imguri with the lazy_tensor type (example tasks, alexnet learner)
• Fix substitution bug
• DataDescriptor should be R6 for more efficient serialization of lazy_tensor
• fully remove the dollar method for the lazy tensor data type
• remove attribute "hash" in c() function at the end
• remove rlang, vctrs, fs
• Move torchvision to suggests
• Don't include batch dim when returning lazy tensor as a list
• address open TODOs in code
• maybe refactor materialize_internal to make it better understandable and fix the bug
• Fix dataset_ltnsr
• Customize hash input for lazy tensor in backend

[sebffischer]

fix stuff