GpuBatchedDot: remove CNDA_THREAD_SYNC from streams implementation as it always syncs, introduce it in the cublasSgemmBatched implementation

Correct In parameter name following Param removal

Fixes the multinomial grad dtype bug

Don't use CpuCorrMM when there is no CPU BLAS

change shape dimensions int32 to int64

Exponential Linear Units (ELUs)

Correct warning re: cuDNN and pool_2d's ignore_border argument

Fix CudaNdarray repr in python 3 and a crash.

Sample multinomial with replacement

fix gh-4036. We need the shape and we don't compute it. It wasn't timmed, so don't try this optimization.

Cleaunup profileMode deprecation in docs

(#4024) cutting the *.a extension in ldlibrary file

Cudnn version, print it, warn when too new, make mandatory in the new back-end.