Raise NotImplementedError at conversion time for reflect/replicate pad >2D (fixes #2576, #2571)

[LeSingh1]

Summary

mb.pad only supports reflect / replicate modes when at most the final two dimensions carry non-zero padding. The generic pad torch converter forwards the mode through unconditionally, so torch.export-decomposed torch.nn.ReflectionPad3d / torch.nn.ReplicationPad3d (which decompose to aten.pad with mode="reflect" / "replicate" and 3 padded dims) currently convert successfully and then explode at predict() time with:

RuntimeError: Error compiling model: "Failed to parse the model specification.
Error: Unable to parse ML Program: in operation pad_cast_fp16:
Padding for more than two dimensions only supports `constant` mode".

The two original issues (#2576 ReflectionPad3D, #2571 ReplicationPad3d) reported this exact failure mode and asked for an ahead-of-time error instead.

Fix

Added a static check inside _translate_torch_args in the pad op converter (coremltools/converters/mil/frontend/torch/ops.py). When mode in ("reflect", "replicate"), the pad list is fully static, and more than two dimensions have non-zero pad, raise NotImplementedError with a message that names the mode, the padded-dim count, and the typical culprit modules.

The check is scoped tightly:

• only static-list pad (dynamic-padding path was already raising for other reasons);
• only reflect / replicate modes (constant and circular unaffected);
• only triggers when padded_dims > 2 — valid 2D cases (ReflectionPad2d, etc.) still convert unchanged.

Tests

Added TestPad::test_pad_reflect_replicate_3d_raises parameterized over both torch.nn.ReflectionPad3d and torch.nn.ReplicationPad3d:

tests/...::test_pad_reflect_replicate_3d_raises[mode=reflect-...]   PASSED
tests/...::test_pad_reflect_replicate_3d_raises[mode=replicate-...] PASSED
======================== 2 passed in 0.64s =========================

I also manually verified torch.nn.ReflectionPad2d(padding=2) on a rank-4 input still passes the new guard and proceeds to the rest of the converter.

Issues

Fixes #2576
Fixes #2571