[QNN EP] Fix quantized model graph I/O name mismatch

[qti-yuduo]

Preserve original ONNX graph I/O names in DLC when offload_graph_io_quantization=1

When offload_graph_io_quantization=1, Q/DQ nodes at graph boundaries are
filtered out, causing fused node I/O names to differ from original ONNX
graph I/O names. This change maps them back so DLC files use user-expected
tensor names.

Changes:

• Add BuildQuantizedIoNameMap() in qnn_model.cc to trace Q/DQ nodes and
  build fused-to-original name mappings
• Add tensor_name_override_ to QnnTensorWrapper for overriding QNN tensor
  name without changing ORT lookup name
• Apply overrides in QnnModelWrapper::AddTensorWrapper() for graph I/O tensors
• Add QuantizedGraphInOutNamesPreserved test