Update PyTorch export tutorial

[pavithraes]

This PR updates the "Exporting StableHLO from PyTorch" tutorial. Specifically:

• Update dependencies, combine the tensorflow dependency in the first section to avoid warnings (about using tensorflow-cpu) on Kaggle
• Minor updates to narration and links, along with an introduction statement
• [TODO] Fix dynamic batch size code, currently commented

(Supersedes #2616, same updates but we get a rich diff here)

[GleasonK]

TorchXLA coverage of dynamic shapes isn't great, for dynamic shape / export workflows we actually recommend torch_xla2 which uses PT->JAX->StableHLO, and has much better coverage for dynamic shapes:
https://github.com/pytorch/xla/tree/master/experimental/torch_xla2

Perhaps we just make a note of "For dynamic shape export we recommend using torch_xla2 (link), which converts PyTorch models to JAX prior to exporting to StableHLO and has high opset coverage along with great dynamic shape support."

If we want to add a tutorial on using torch_xla2, it likely wouldn't be too hard, here's a recent example (discord invite needed to view):
https://discord.com/channels/999073994483433573/1281512734902714379/1298361762839527467

I believe the following should work:

# Install torch_xla2
pip install torch-xla2 # Note: may need to follow GH instructions linked above but pypi package looks up to date.

# Export the program from native pytorch using torch.export
from torch.export import Dim
batch = Dim("batch", min=4, max=6)
dynamic_shapes = ({0: batch},)
dynamic_export = export(resnet18, sample_input, dynamic_shapes=dynamic_shapes)

# Convert the exported program to StableHLO using torch_xla2
from torch_xla2 import export
stablehlo = export.exported_program_to_stablehlo(dynamic_export).mlir_module()
print(stablehlo)

Note the print will include lots of debug info, so maybe use some indexing, or print starting with the word @main for 1k characters or something to keep the tutorial readable. I.e. print(stablehlo[1000:2000]) or something.