Fix PyTorch 2.13 named overload compatibility

[juicyzhou]

Fixes #101.

Summary

• add a small compatibility helper for looking up optional ATen overloads
• register named-tensor autocast policies and decompositions only when the installed PyTorch still provides them
• add regression coverage for importing torchax and constructing the default environment
• add a focused PyTorch 2.13 compatibility CI job while keeping the existing PyTorch 2.8 full test matrix unchanged

Root cause

PyTorch 2.13 removed named tensors and their ATen overloads. torchax.amp accessed two removed overloads at module import time, and torchax.decompositions referenced another 20 removed overloads while constructing the default environment. Accessing any missing overload raises AttributeError before torchax can be used.

The compatibility lookup preserves the existing registrations on older PyTorch versions and skips only unavailable named-tensor overloads on 2.13 and later.

Validation

• PyTorch 2.8.0, Python 3.12: pytest -q test/test_torch_compat.py test/test_amp.py — 6 passed
• PyTorch 2.13.0, Python 3.12: pytest -q test/test_torch_compat.py test/test_amp.py — 6 passed
• PyTorch 2.13.0: import torchax and torchax.default_env() succeed
• ruff check and ruff format --check pass for all changed Python files

[google-cla]

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.

[qihqi]

no need this hack. Just deleting in decompositions would be good.