Support variable length inputs in RNNs

[unnonouno]

cuDNN's RNN library supports variable length inputs. The RNN functions such as LSTM and GRU in chainer does not support that in native (users can support this feature with where function). Its specification is simple and useful, and it is required to keep compatibility with cuDNN's.

I write the detail below:

For example there are three input sequences, a, b and c:

• a = [a1, a2, a3, a4]
• b = [b1, b2, b3]
• c = [c1]

where each element is a vector. These sequences need to be sorted by descending length.
A RNN function gets four variables, x1, x2, x3 and x4:

• x1 = [a1, b1, c1]
• x2 = [a2, b2]
• x3 = [a3, b3]
• x4 = [a4]

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed after 30 days if no further activity occurs. Thank you for your contributions.

[stale]

This issue is closed as announced. Feel free to re-open it if needed.