WIP: TensorParallel with new strategy

[awaelchli]

Shows how the new ModelParallelStrategy could be applied in generate/tp.py.

Some caveats:

• Need to apply quantization in the parallelize function (to be done)
• DTensors don't work with torch.inference_mode

There is room to apply more parallelism. For example, the RMSNorm could leverage SequenceParallel.