Fix torch gelu converter ignoring the approximate keyword argument

[Yigtwxx]

Summary

The torch gelu converter only reads approximate when it arrives as a positional input, which is how TorchScript serializes it. torch.export and the ExecuTorch edge dialect keep it as a keyword argument (aten.gelu.default(x, approximate='tanh')), so nn.GELU(approximate="tanh") silently converts to the exact gelu on those frontends.

This affects every model that uses the tanh variant, most visibly the Hugging Face gelu_pytorch_tanh activation (ACT2FN["gelu_pytorch_tanh"] wraps nn.functional.gelu(approximate="tanh")), which is the default activation of the Gemma / Gemma 2 / Gemma 3 configs and of SigLIP and RecurrentGemma among others. The two gelu variants differ by up to ~5e-4 per activation, and the existing test_gelu runs with atol=1e-3, which is why the mismatch never showed up.

Minimal reproduction on main:

import torch, coremltools as ct

model = torch.nn.GELU(approximate="tanh").eval()
x = torch.randn(2, 8)
ep = torch.export.export(model, (x,)).run_decompositions({})
prog = ct.convert(ep, inputs=[ct.TensorType(shape=x.shape)], convert_to="milinternal")
print(prog.functions["main"].find_ops(op_type="gelu")[0].mode.val)
# main: EXACT        (torch.jit.trace of the same module gives TANH_APPROXIMATION)
# this PR: TANH_APPROXIMATION

Implementation

• gelu now falls back to _get_kwinputs(context, node, "approximate") when the argument is not positional, unwraps a Var, and maps "tanh" to TANH_APPROXIMATION. Any value other than "none" / "tanh" raises a ValueError instead of tripping an assertion.
• "none" and "tanh" are added to TORCH_STRING_ARGS so the export frontend binds them without the "neither a name of existing var nor a torch string argument" warning.

TorchScript behaviour is unchanged.

Tests

test_gelu keeps its tolerance but now also asserts the mode of the emitted gelu op (EXACT vs TANH_APPROXIMATION) for every frontend. On main the new assertion fails for the TORCHEXPORT and EXECUTORCH approximate="tanh" cases and passes for TorchScript; with this change all 45 test_gelu cases pass, and TestActivation shows no new failures compared to main.