[MXNET-1327] Allow RNN Layers to be initialized to fp16

[ThomasDelteil]

Description

Currently if you hybridize a RNN layers, it doesn't work with fp16 except if you explicitly give fp16 states as initial states. This makes it cumbersome to use.
This PR updates the RNN layers to have a dtype that is used for the default states and is updated upon casting.

Currently

net = gluon.rnn.LSTM(5, bidirectional=True)
net.collect_params().initialize(ctx=mx.gpu())
net.cast('float16')
net.hybridize()
net(mx.nd.ones((1,5,5), dtype='float16', ctx=mx.gpu()))

gives

MXNetError: Error in operator lstm5_rnn0: [01:23:01] src/operator/./rnn-inl.h:750: Check failed: (*in_type)[i] == dtype (0 vs. 2) This layer requires uniform type. Expected 'float16' v.s. given 'float32' at 'state'

Stack trace returned 10 entries:
[bt] (0) /home/ubuntu/anaconda3/lib/python3.6/site-packages/mxnet/libmxnet.so(+0x40e86a) [0x7fd8952bf86a]
[bt] (1) /home/ubuntu/anaconda3/lib/python3.6/site-packages/mxnet/libmxnet.so(+0x40eea1) [0x7fd8952bfea1]

After the fix

[[[ 0.001968  0.002743 -0.03683   0.015366 -0.02847   0.04007   0.00813
    0.002373  0.014656  0.03268 ]
  [ 0.001968  0.002743 -0.03683   0.015366 -0.02847   0.04007   0.00813
    0.002373  0.014656  0.03268 ]
  [ 0.001968  0.002743 -0.03683   0.015366 -0.02847   0.04007   0.00813
    0.002373  0.014656  0.03268 ]
  [ 0.001968  0.002743 -0.03683   0.015366 -0.02847   0.04007   0.00813
    0.002373  0.014656  0.03268 ]
  [ 0.001968  0.002743 -0.03683   0.015366 -0.02847   0.04007   0.00813
    0.002373  0.014656  0.03268 ]]]
<NDArray 1x5x10 @gpu(0)>

Checklist

Essentials

Please feel free to remove inapplicable items for your PR.

• The PR title starts with [MXNET-$JIRA_ID], where $JIRA_ID refers to the relevant JIRA issue created (except PRs with tiny changes)
• Changes are complete (i.e. I finished coding on this PR)
• All changes have test coverage:
• Unit tests are added for small changes to verify correctness (e.g. adding a new operator)
• Nightly tests are added for complicated/long-running ones (e.g. changing distributed kvstore)
• Build tests will be added for build configuration changes (e.g. adding a new build option with NCCL)
• Code is well-documented:
• For user-facing API changes, API doc string has been updated.
• Check the API doc at http://mxnet-ci-doc.s3-accelerate.dualstack.amazonaws.com/PR-$PR_ID/$BUILD_ID/index.html
• To the my best knowledge, examples are either not affected by this change, or have been fixed to be compatible with this change

Changes

• Added a non-breaking dtype param to RNN layers

[haojin2]

I think you've linked with the wrong JIRA ticket for this PR... the title of MXNET-939 JIRA ticket is actually "Variance Operator".

[roywei]

Thanks for the contribution, LGTM!

[roywei]

@mxnet-label-bot add[Gluon, RNN, FP16]

[eric-haibin-lin]

Could you fix the test? Thanks!