Potentially redundant bias in convLSTM

[lcnature]

https://github.com/deepmind/sonnet/blob/7da90b8e63fb684b91a12e1e9a42999962804e3a/sonnet/python/modules/gated_rnn.py#L1129

Hi, I realized that the variable of convLSTM includes two biases if use_bias is set to true. And I think it is because here the convolution is performed separately on input and state. But effectively only one set of bias is necessary. Perhaps adding bias at the end instead of within the convolution as done in the tensorflow contrib version (https://github.com/tensorflow/tensorflow/pull/8891/files) would be better.

[malcolmreynolds]

Fixed as of e15eb74 - the old behaviour is still the default as we want to maintain checkpoint compatibility, but setting legacy_bias_behaviour to False will apply only one bias.