Add support for non-zero padding

[lgeiger]

This adds support for non zero padding for QuantConv1D, QuantConv2D, QuantConv3D, QuantDepthwiseConv2D, QuantSeparableConv1D and QuantSeparableConv2D in order to better support larq/compute-engine#252

We can add support for locally connected layers and transposed convolutions in the future.

[lgeiger]

@larq/core how do you feel about pad_values vs padding_values?

[AdamHillier]

I'm fine with either, no particular preference.

[leonoverweel]

From an initial look this LGTM, but given how complex it is I think another few sets of eyes would be useful. CC @Tombana, given your experience with padding, can you take a look?

[Tombana]

Looks good to me. I'd say pad_value (singular) makes slightly more sense but since the tf op also uses it in plural form I think this is good.

[koenhelwegen]

Looks good! Only thing is I would prefer to test whether the layer actually uses non-zero padding (if I'm correct the test would now pass even if _is_native_padding() always returns True)

[lgeiger]

Only thing is I would prefer to test whether the layer actually uses non-zero padding (if I'm correct the test would now pass even if _is_native_padding() always returns True)

Good point. I added a check to make sure tf.pad gets called with the correct values in f03c714