Fix Sparse grad crash and implement mixed dtype in sparse Mul/Add

I changed how format is specified. There was no bug as the input
variable are put in the key. Now I make explicit in the op key the
format.

This simplify at a few places the code too.

Doc

Added numpy wrapper for poisson distribution

Fix Gpu crash on windows

Fix crash during unpickling of Theano variable

Hide symbols in compiled modules

Minor fixes

Python3 compatibility in test program compilation

Similar fix to #1660.

Without this fix, a `'str' does not support the buffer interface` error is raised in Python 3.3 on Mac OSX 10.9. One place it occurs is during testing for the Mac OS sdot bug (line 73, blas_headers.py)

Faster opt

When an apply node execution raise an error, print the inputs values when they are scalars.

new Gpu reduce based on the version in Theano that is much faster.

Rename GpuCAReduce to GpuCAReduceCuda and reuse the name GpuCAReduce for GpuCAReduceCPY to allow reloading older pickled files.