Create `LearnerTorchFTTransformer`

[sebffischer]

Before implementing the learner. Just build it as a Graph.
The graph consists of:

1. Two input nodes (one for tokenization of categoricals, one for numerics). Need to maybe handle what happens when only numerics or only categoricals exist, but ignore for now.
2. Apply the Categorical encoder to categorical inputs, numeric encoder to numerical inputs.
3. A Merge PipeOp that combines the categoricals and numerics
4. A PipeOpTorchBlock that takes in the PipeOpTorch that represents a single transformer block (the PipeOpTorchBlock is responsible for repeating it multiple times
5. Build the output head using existing PipeOpTorch.