apply_weight_drop required changes to integrate mxnet 1.3

[szha]

our existing implementation on apply_weight_drop relied on a non-public attribute in gluon RNN blocks that were removed apache/mxnet#11482 when making the RNN blocks hybrid, and it does not have test, which will prevent us from integrating with mxnet 1.3. We need:

• fix apply_weight_drop and upgrade the mxnet dependency to allow 1.3 integration.
• add test for apply_weight_drop
• bump up minor version to 0.4 which relies on 1.3.