train_defeat_zerglings uses wrong dimensions for q_func

[rwill128]

I am not sure how to best address this, but it looks like train_defeat_zerglings (and potentially the Q learning version of train_mineral_shards) are not operational with TF 1.14.

The error is thrown in line 1024 of layers.py

    if conv_dims is not None and conv_dims + 2 != input_rank:
      raise ValueError('Convolution expects input with rank %d, got %d' %
                       (conv_dims + 2, input_rank))