[PyTorch] Support divisor_override for avg_pool2d and avg_pool3d #2745

Description

The PyTorch frontend currently rejects avg_pool2d and avg_pool3d when divisor_override is provided.

Minimal reproduction

import torch
import coremltools as ct

class Model(torch.nn.Module):
    def forward(self, x):
        return torch.nn.functional.avg_pool2d(
            x,
            kernel_size=2,
            divisor_override=3,
        )

model = Model().eval()
example = torch.rand(1, 1, 4, 4)
traced = torch.jit.trace(model, example)

ct.convert(
    traced,
    inputs=[ct.TensorType(shape=example.shape)],
    convert_to="mlprogram",
)

Current behavior

Conversion fails with:

ValueError: divisor_override is not supported for avg_pool2d

Expected behavior

Models using a static, nonzero divisor_override should convert successfully when the operation can be represented using MIL operations, while preserving numerical parity with PyTorch.

Possible implementation

For configurations where average pooling uses a fixed full-kernel divisor:

1. Lower the operation using the existing average-pooling implementation.
2. Multiply the result by:

number_of_kernel_elements / divisor_override

An initial implementation could support:

• avg_pool2d and avg_pool3d
• Static kernel dimensions
• Positive constant divisor_override
• Zero padding, or count_include_pad=True

Cases combining nonzero padding with count_include_pad=False may require separate handling because the effective divisor can vary near boundaries.

I would be happy to implement this and add regression tests if this scope is acceptable.

Environment

• PyTorch: 2.8.0
• coremltools: 9.0
• Platform: macOS / Apple silicon