Improve convolution (#5284)

* Add im2col for convolution

* Implemented Forward function for convolution op

* Added req type to col2im functions

* Added backward for computing gradients wrt image data

* Added backward for weights and bias

* Fixed bugs and added one test case

* Added one test case

* Added infer shape for 1d conv

* Fixed a bug for computing gradient of bias

* Removed Convolution2

* Renamed old and new convolution implementations

* Fixed lint errors

* Fixed a lint error

* Pass stream to im2col gpu functions

* Fixed convolution n-D gpu bugs

* Added 1/2/3-D cpu/gpu convolution test cases

* Move im2col.cc and im2col.cu to header files

* Consolidated interfaces of im2col and col2im for cpu and gpu

* Clean up and add test cases

* Activated test case

* Fixed a typo

* Enable openMP for 2d image cpu convolution

* Reverted openmp

* Added mshadow change

* Fixed warnings of ConvertLayout in mshadow

* Fixed unsigned/signed integer comparison

* Fixed mkl build error on jenkins