Fixes needed for VQVAE

[marksgraham]

I'm training some VQ-VAEs and have started to noticed a few minor issues. I plan to keep track of them here as I find them, with the intention of fixing them all in one PR later on.

• These argument docstrings imply the args should be ints but they are actually sequences:

num_channels: number of channels at the deepest level, besides that is num_channels//2 . Defaults to 192.
num_res_channels: number of channels in the residual layers. Defaults to 64.

• The docstrings for all the arguments state the default values. This isn't keeping with the style of the rest of the package, and makes the code harder to maintain as changes to default in the function arguments need to be reflected in the dosctrings too.

• This error on L166 repeates num_res_channels twice, the first one should be num_channels:

    f"downsample_parameters, upsample_parameters, num_channels and num_res_channels must have the same number of"
    f" elements as num_levels. But got {len(downsample_parameters)}, {len(upsample_parameters)}, "
    f"{len(num_res_channels)} and {len(num_res_channels)} instead of {num_levels}."