Add support for Python 3.11

[nblomerus]

Summary

Adds Python 3.11 to TFX's supported and tested matrix. Closes #7743.

All 12 CI jobs pass on the contributor's fork (3 Python versions × not-e2e/e2e × DEFAULT/NIGHTLY dependency selectors): https://github.com/nblomerus/tfx/actions/runs/26000991135

Changes

Packaging

• requires-python upper bound bumped from <3.11 to <3.12 in the three pyproject.toml files (top-level + package_build/tfx + package_build/ml-pipelines-sdk).
• Programming Language :: Python :: 3.11 classifier added in each.

CI matrix

• '3.11' added to ci-test.yml; 'cp311' added to wheels.yml.
• strategy.fail-fast: false added so individual job failures don't cancel the rest of the matrix, and was essential for debugging this PR. Happy to split this into a separate PR if preferred.

Dependency pins (test_constraints.txt + nightly_test_constraints.txt)

The TFX 1.17 cp311 wheels require newer versions of several deps than the cp39/cp310 wheels, so this uses PEP 508 environment markers to pin per-Python-version. Each == pin is conditional on Python version where the existing pin would conflict with the cp311 wheels' own Requires-Dist:

Package	<3.11	>=3.11	Why
apache-beam	2.50.0	2.53.0	tfx-bsl/tfdv cp311 wheels require beam >=2.53
protobuf	4.21.12	4.25.5	tfx-bsl/tfdv/ml-metadata cp311 wheels require >=4.25.2; beam 2.53 requires !=4.21.*
pyarrow	(unconstrained)	10.0.1	10.0.0 has no cp311 wheel and fails sdist build under setuptools 70
tensorflow-io-gcs-filesystem	0.24.0	0.37.1	0.24.0 has no cp311 wheel

Plus one unconditional bump: grpcio-status 1.48.2 → 1.62.3. google-api-core requires grpcio-status>=1.49.1 on Python 3.11+; grpcio-status 1.63+ would require protobuf 5, conflicting with TFX's protobuf<5. 1.62.3 is the last release using protobuf 4.

Test fixes

Two test files had patterns that broke on Python 3.11:

• tfx/components/infra_validator/request_builder_test.py — mock.patch.object(..., wraps=Cls) returns a _SentinelObject for .return_value until the wrapped class is called (Python 3.11 mock behaviour change). Switched to autospec=True. autospec is strict about kwargs so the examples= arg is passed by keyword.
• tfx/components/trainer/rewriting/tflite_rewriter_test.py — mock.patch('tensorflow.lite.TFLiteConverter.from_saved_model') fails on 3.11 because mock's string-path resolution doesn't trigger TF's lazy tf.lite module wrapper. Switched to mock.patch.object(tf.lite.TFLiteConverter, ...), which references the object directly. This keeps TFLite functional — a deliberately different approach than Adding python 3.11 support and Removing TF-Lite Dependency #7764, which gutted TFLite to side-step the issue.

Scope choices

• 3.11 only, not 3.12. The TFX-ecosystem cp312 wheels (tfdv, tfx-bsl, ml-metadata) aren't published on PyPI or the nightly index yet, so 3.12 support is blocked on Google releasing those. build: enable Python 3.12 support and relax core dependency bounds #7821 covers the broader 3.12 work (via local wheel builds).
• Conditional pins, not range widening. I considered loosening == pins to ranges (>=) and letting pip's resolver pick. But the current constraints file is a frozen snapshot for reproducibility, and the maintainers can tell me if they'd prefer the range approach.

Test plan

• All 12 CI jobs green on contributor fork.
• Each modified test (request_builder_test.py::testBuildRequests_TFServing, ::testBuildRequests_DefaultArgument, tflite_rewriter_test.py::testInvokeTFLiteRewriterQuantizationFullIntegerFailsNoData, ::testInvokeTFLiteRewriterWithSignatureKey) passes on 3.9, 3.10, 3.11.
• No tests skipped or xfailed as a result of this change.