* The original SparseRowMatrix use two fields to store each rows,
  which let code very confusing. Try to extract a RowBuffer class,
  for SparseRowMatrix data storage, and manage auto-growth logic.

* Also fix a protential bug in WarpCTCLayer.cpp

Also fix compile issues

* But not reproduce the problem.

When the define_py_data_sources2 has both train_list and test_list,
for job=test, the trainer will create both dataProvider_ and testDataProvider_.
But dataProvider_ is not used. This causes SIGSEGV at finishAsync() because asyncLoader_ is not created.

Change-Id: If579f715f80a70ebc795094792c3436bfa0f5746

Added a test in test_NetworkCompare to verify cudnn and exconv, fixed a bug in cudnn_conv in dealing with groups

* Different Type compare.
* ostream << should pass a const object.
* remove always true checks.

* Different Type compare.
* ostream << should pass a const object.
* remove always true checks.

Change-Id: Ic80845c892c96c37a0df0ddc433fe1aeaa5a9d1c

out = w * input
where w is a parameter of size 1

Change-Id: Ife682d62323ceb1a20cbbf6269421b20a862d888

* Check all files by pre commit hooks