Fix ONNX optional input handling by preserving positional indices in getNodeInputs()

[Copilot]

getNodeInputs() compacted the returned vector by skipping empty input names, breaking positional indexing for any op with optional inputs. For Resize v11, this caused the ROI input (index 1, commonly empty) to shift scales and sizes to wrong indices, and incorrectly rejected valid nodes.

Changes

• getNodeInputs() refactor — returns nullptr for empty input names instead of skipping them; inputs[N] now always refers to the Nth ONNX input regardless of preceding optional inputs
• Resize.cpp convertResizeV11 — drops the onnx_node.input() workaround; uses getNodeInputs() directly with clean positional access:
  • inputs[0] → X, inputs[1] → ROI (nullptr when absent), inputs[2] → scales, inputs[3] → sizes
• Conv.cpp / ConvTranspose.cpp — bias detection updated from inputs.size() > 2 to inputs.size() > 2 && inputs[2] != nullptr to correctly handle a bias with an empty name
• Test — added ConverterContextTest/getNodeInputs_preserves_positional_indices verifying that a node with inputs ["a", "", "b"] yields [a, nullptr, b]

// Before: skipped empty names → compacted vector
// ["X", "", "scales"]  →  [x, scales]   // inputs[1] is wrongly scales

// After: nullptr for absent inputs → positional vector preserved
// ["X", "", "scales"]  →  [x, nullptr, scales]   // inputs[2] is correctly scales

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.