Support return_indices for max_pool2d in the torch frontend

[adityasingh2400]

Fixes #2456.

Problem

max_pool2d with return_indices=True could not be converted. The torch frontend bound the indices output to None, so any model that read the indices failed with ValueError: Torch var <n> not found in context:

class Model(torch.nn.Module):
    def forward(self, image):
        _, ixs = F.max_pool2d(image, kernel_size=3, stride=1, padding=1, return_indices=True)
        return ixs

ct.convert(torch.jit.trace(Model().eval(), torch.rand(1, 3, 224, 224)),
           inputs=[ct.TensorType(shape=(1, 3, 224, 224))])

Fix

MIL has no max-pool-with-indices op, so the indices are recovered from the pooling windows. The input is padded so window extraction lines up with the pooled output, the windows are materialized with sliding_windows (batch and channel are folded together to stay within Core ML's rank-5 limit), reduce_argmax finds the selected element inside each window, and the within-window position is mapped back to a flattened input coordinate, which is what PyTorch returns.

A few details match PyTorch exactly:

• The pooled output's spatial size is used as the source of truth, so the ceil_mode rule that drops a trailing window starting inside the bottom/right padding is honored.
• reduce_argmax returns the first occurrence on ties, matching PyTorch's tie-breaking.
• Padded cells are filled with the lowest representable float so they are never selected, matching PyTorch treating padding as -inf.

The values output is unchanged. 1D/3D return_indices and dynamic input shapes raise a clear NotImplementedError instead of silently producing wrong results.

Testing

Added test_max_pool2d_return_indices, parametrized over kernel size, stride, padding, and ceil_mode for both the TorchScript and torch.export frontends. Indices are compared against the torch reference on the fp32 path; a fp16 backend can rank two near-equal window elements differently from torch's fp32 argmax, so that combination is skipped for the index comparison. The existing max_pool tests continue to pass.