* Compose GRU step naive layer in trainer config helpers.
  * It is uses mixed_layer for gate.
  * It supports ERROR_CLIPPING, DROPOUT
* Add error clipping in MT demo.
* Fix #1143
* Fix #1891

It should be always equals to label's size.

The reason for adding the function is exemplified in the following hypothetical code:

mem = memory(name=None, size=256)
hidden = fc_layer(input=mem)
state = hidden + x
mem.set_input(state)

In the above code segment, it would be very annoying if we require the user to provide
the name at memory() call because the layer name of state is automatically generated
and is not easy to set it.

Change-Id: I918bf1d3d5c26addd88a6f7021e98b3e0e9df494

* It seems some other change breaks the async load. It will cause data
  error in GPU mode.