Fix bert nlp models failing input generation by overriding `token_type_ids` to limit its range to [0,2) and some refactoring

[renxida]

Note: this causes each nlp model to create an additional onnx inference session. A larger scale refactoring may be required to fix this.

Testing: I ran this on local with a bert model and was able to get to a numeric mismatch.

Full list of changes:

1. Refactor input generation for ONNX models:

   • Split get_node_shape_from_dim_param_dict() out of generate_input_from_node() for better modularity.
   • Improve docstrings for better function descriptions.

2. Enhance type annotations:

   • Add return type hints to construct_inputs() and get_sample_inputs_for_onnx_model().
   • Use TestTensors type consistently for input/output tensors.

3. Fix NLP(Bert)-specific input generation:

   • Override construct_inputs() in NLP models to handle 'token_type_ids' correctly.