[Data] Remove tfx-bsl support from read_tfrecords

[goutamvenkat-anyscale]

Summary

Ray Data now pins pyarrow>=17, but tfx-bsl caps pyarrow<11, so the optional tfx-bsl-backed reader for read_tfrecords can no longer install alongside Ray. Remove the tfx-bsl path entirely; TFRecord files continue to read via the default pure-pyarrow + tensorflow path.

Changes

• Drop TFXReadOptions and the tfx_read_options kwarg on ray.data.read_tfrecords; drop the public re-export.
• Remove _tfx_read_stream, _resolve_full_path, and the schema-inference helpers used only by the tfx-bsl path.
• Delete the dedicated CI: datatfxbslbuild docker image, data_tfxbsl_depset / relaxed_data_tfxbsl_ci_depset depset entries and lockfiles, the "TFRecords (tfx-bsl) tests" Buildkite step, python/requirements/data/tfxbsl.txt, and the tfxbsl test tag (so test_tfrecords now runs in the standard arrow v17/v23 data jobs).
• Drop the tfx_bsl.* pyrefly ignore, the TFXReadOptions doc entry, and the corresponding pydoclint baseline line.

read_tfrecords and TFXReadOptions were both @PublicAPI(stability=\"alpha\"), so per Ray's API stability policy breaking changes are permitted without a deprecation cycle.

[gemini-code-assist]

Code Review

This pull request removes the tfx-bsl dependency and all associated logic for reading TFRecord files, including the TFXReadOptions class and specialized schema inference. The read_tfrecords API is simplified to rely solely on the standard TensorFlow-based implementation. Review feedback identifies that the new function signature for read_tfrecords and the instantiation of TFRecordDatasource are missing **kwargs, which prevents users from passing file-based configuration options such as compression. Furthermore, it is suggested to retain the GZIP test cases to verify that compression support continues to work correctly in the remaining implementation.

[gemini-code-assist]

The read_tfrecords function signature is missing **kwargs. This prevents users from passing additional file-based datasource options like arrow_open_stream_args (used for compression) to the underlying TFRecordDatasource. Since the tests in test_tfrecords.py attempt to use this argument, it should be added to the signature to maintain consistency with other read_* APIs and support file-based options.

    file_extensions: Optional[List[str]] = None,
    concurrency: Optional[int] = None,
    override_num_blocks: Optional[int] = None,
    **kwargs,
) -> Dataset:

[gemini-code-assist]

The TFRecordDatasource constructor should receive the **kwargs from read_tfrecords to ensure that file-based options (like compression settings) are correctly passed to the datasource.

        shuffle=shuffle,
        include_paths=include_paths,
        file_extensions=file_extensions,
        **kwargs,
    )

[gemini-code-assist]

The parametrization for test_read_tfrecords has been reduced to only test with compression=None. Since the default path (pure-pyarrow + tensorflow) supports reading compressed TFRecord files via arrow_open_stream_args, we should retain the GZIP test cases to ensure continued support and avoid regressions in the remaining implementation.

    "with_tf_schema,compression",
    [
        (True, None),
        (True, "GZIP"),
        (False, None),
        (False, "GZIP"),
    ],