Import ONNX Pad including negative-padding crop semantics

[matthargett]

ONNX Pad is unsupported at 0bcd517, including a crop used to retain convolution history in this small Mamba fixture.

• Minimal reproducer: opset 13, FP32 input [1,2,10], constant INT64 pads [0,0,-6,0,0,0], constant mode, omitted constant_value.
  • Expected output: [1,2,4], exactly the last four values along the final axis. ONNX checker and ONNX Runtime CPU agree.
  • convert-onnx --input crop.onnx --inline-weights fails with unsupported operator: Pad.
• The model's negative padding is cropping, not an error or zero padding. WebNN's nonnegative padding arguments therefore need a slice composition; adding only positive pad support would not unblock this export.

Suggested acceptance checklist:

• Resolve constant pads, optional scalar value and opset-18 axes; normalize axes and validate ranks/ranges.
• Lower cropping to slice, then add any positive padding; cover pure crop and mixed crop/pad.
• Map constant/edge/reflect semantics, including ONNX reflect to WebNN reflection.
• Reject unresolved runtime parameters explicitly; do not substitute maximum dimensions. Dynamic producer provenance is tracked separately in Preserve producer-based provenance for dynamic Slice and Range bounds #20.
• Test shapes and values against ONNX Runtime, including omitted optional inputs, and verify emitted JS uses WebNN's positional padding arguments.

This is an importer/decomposition gap, not a request to change WebNN pad. The fixture also has a separate hard-unrolled sequence-length restriction, so importing it will not establish general dynamic decode correctness.

Minimal fixture (Python onnx):

import onnx
from onnx import helper as h, TensorProto as T
x = h.make_tensor_value_info("x", T.FLOAT, [1, 2, 10])
y = h.make_tensor_value_info("y", T.FLOAT, [1, 2, 4])
pads = h.make_tensor("pads", T.INT64, [6], [0, 0, -6, 0, 0, 0])
graph = h.make_graph([h.make_node("Pad", ["x", "pads"], ["y"])], "crop", [x], [y], [pads])
model = h.make_model(graph, opset_imports=[h.make_opsetid("", 13)], ir_version=8)
onnx.checker.check_model(model)
onnx.save(model, "crop.onnx")