Handle Keras 3 tensor metadata in TF2 models

[kiwigitops]

Summary

• read TF2 tensor dtype and shape from type_spec when available, with a Keras 3 fallback to dtype and shape
• normalize tensor metadata through TensorFlow before building ResponseInfo
• use layer metadata when Keras 3 output names are generic, so Conv2D responses still classify correctly

Fixes #2

Tests

• $env:PYTHONPATH='src/dnikit;src/dnikit_tensorflow'; python -m pytest -o addopts= src/dnikit_tensorflow/tests/test_tf2_model_loaders.py::test_get_tensor_dtype_and_shape_without_type_spec src/dnikit_tensorflow/tests/test_tf2_model_loaders.py::test_convert_tf_operation_uses_layer_metadata_when_output_name_is_generic src/dnikit_tensorflow/tests/test_tf2_model_loaders.py::test_tf_load_from_memory -q
• python -m compileall -q src/dnikit_tensorflow/dnikit_tensorflow/_tensorflow/_tf2_model.py src/dnikit_tensorflow/tests/test_tf2_model_loaders.py
• git diff --check