Add bool dtype in scalar and tensor.

Fixed typo in config doc

make var() return float16 when input is float16.

Insert GpuSolve only when CULA is available

Doc update about cudnn and cudnn RNN.

Fix a problem where the samples buffer would not be filled.

Sum with a complex input and scalar output doesn't work because of the
ambiguity of casting complex to scalar.  The reverse doesn't work
because of the ambiguity of the casting in the gradient.  Therefore
disallow both.

New optimization: DimShuffle(Alloc(...)) -> Alloc(..., use dims of 1 for new dimensions) #5024

Log sum exp optimization for numerical stability

Manual revert of 29e67107