Preserve tensor feature padding_value during serialization

[this-AR]

Summary

This change ensures the padding_value of tensor features is included during schema serialization and restored when loading. This fixes losing custom padding values (e.g. -1) when saving and loading sequential datasets or tokenizers.

Files changed

• schema.py — include padding_value in _get_object_args and ensure _create_object_by_args passes it to TensorFeatureInfo.
• padding_value_serialization_fix.md — describe the bug and fix.
• test_schema.py — add/verify test_tensor_schema_serialization_preserves_padding_value.
• test_sequential_dataset.py — verify save/load preserves feature padding values.

Motivation

Previously custom padding values were not saved with tensor schema metadata and were reset to 0 on load. This breaks workflows relying on non-zero padding sentinels.

Backward compatibility

Older saved metadata without padding_value remains compatible because TensorFeatureInfo defaults padding_value to 0.

Tests

pytest -q tests/data/nn/test_schema.py::test_tensor_schema_serialization_preserves_padding_value \
       tests/data/nn/test_sequential_dataset.py::test_save_and_load_sequential_dataset