Raise NotImplementedError for unsupported TF symmetric pad mode

[adityasingh2400]

Issue

Fixes #2280.

When converting a TensorFlow model, both the Pad and MirrorPad op handlers in coremltools/converters/mil/frontend/tensorflow/ops.py silently rewrite TF's symmetric padding mode to MIL's reflect:

mode = node.attr.get("mode", "constant").lower()
if mode == "symmetric":
    mode = "reflect"

The two modes are not equivalent. TF symmetric reflects including the edge value, while reflect excludes it. For input [1, 2, 3, 4] padded by 2 on each side:

• symmetric: [2, 1, | 1, 2, 3, 4 | 4, 3]
• reflect: [3, 2, | 1, 2, 3, 4 | 3, 2]

So a tf.pad(..., mode='SYMMETRIC') currently converts without any warning and the resulting Core ML model computes different, wrong values. This is a silent miscompilation.

Fix

MIL pad supports only constant, reflect, and replicate (see mil/ops/defs/iOS15/tensor_operation.py), so symmetric cannot be expressed directly. As the reporter suggested, the right behavior is to raise a clear error rather than emit a wrong result. Both handlers now raise NotImplementedError naming the op and the unsupported mode. The reflect and constant paths are unchanged.

Testing

Added TestPad::test_symmetric_unsupported, which asserts that converting a tf.pad(mode='SYMMETRIC') graph raises NotImplementedError. Verified locally with TensorFlow 2.19:

• before the fix, the TF -> MIL conversion emits pad(..., mode="reflect") for a SYMMETRIC input (silent wrong mapping confirmed)
• after the fix, SYMMETRIC raises NotImplementedError through both the MirrorPad path (tf.pad lowers SYMMETRIC to MirrorPad) and the Pad handler branch
• REFLECT and CONSTANT still convert correctly and emit mode="reflect" / mode="constant"

The new test passes for both backends. End-to-end runtime comparison was not run here because this environment lacks the compiled Core ML runtime, but the change is confined to the frontend mode mapping.