Candle: better pooling op support

[kingwingfly]

Dear developpers,

I'm trying to train a ResNet model with Candle as its backend.

But it has a runtime panic raised here:

"max_pool2d_with_indices is not supported by Candle"

The model contains MaxPool2d and AdaptiveAvgPool2d, which look not well supported by Candle.

#[derive(Module, Debug)]
pub struct ResNet<B: Backend> {
    ...
    relu: Relu,
    maxpool: MaxPool2d,
    ...
    avgpool: AdaptiveAvgPool2d,
    fc: Linear<B>,
}

I really want to try Candle, but how to avoid MaxPool2d not working bug, or is there any alternative implementation? It is so kind of you to give me any inspiration!