Migrate NLP model inputs to BatchEncode and add named ONNX input binding

[Copilot]

This PR replaces flat tensor-array tokenization outputs with a Hugging Face-style BatchEncode contract and updates ONNX execution to bind inputs by semantic names when available. It also preserves backward compatibility for legacy Tensor<long>[] model pipelines.

• Input contract unification

  • Added BatchEncode in FAI.Core.Pipelines to represent input_ids, optional attention_mask, and optional token_type_ids.
  • Updated tokenizer/tensorization flow to emit BatchEncode directly instead of positional tensor arrays.

• NLP pipeline type migration

  • TextTensorization and TextMultipleChoiceTensorization now output BatchEncode.
  • Downstream pipeline wiring in examples/integration paths was updated to compose ONNX stages with BatchEncode.

• ONNX named-input support

  • ONNX executors now accept IPipeline<BatchEncode, TensorOutputs<float>>.
  • Input resolution maps BatchEncode fields to model-declared names in model order (input_ids, attention_mask, token_type_ids).
  • If the model exposes HF names, only declared inputs are passed; missing required declared inputs fail fast with clear exceptions.
  • Legacy positional Tensor<long>[] execution remains supported through the existing interface path.

• Factory/executor alignment

  • ModelExecutorFactory.CreateModelPipeline(...) now returns IPipeline<BatchEncode, TensorOutputs<float>>.
  • Legacy tensor-array compatibility wrappers remain available for existing callers still using positional inputs.

• Targeted ONNX coverage updates

  • Updated ONNX factory/executor tests to validate BatchEncode execution behavior, including optional-input handling and model-name resolution behavior.

services
    .AddPipeline<ReadOnlyMemory<string>>()
    .Then<ReadOnlyMemory<TokenizedText>, TextTokenization>()
    .Then<BatchEncode, TextTensorization>()
    .ThenOnnxModel();