The optimization before was all or nothing, which is a bad compromise for
scan. This new optimization tries to work in place on each outputs, and
keeps only those for which it can. This way you do not get your entire op to
be non-inplace because of a single output that can not be computed inplace.

Add MergeOptimizer in canonicalize phase

Fix advanced indexing with constants (not variables)

Closes gh-607

fix test in float32, but also check float64 at the same time.

Disable ConvOp grad when subsampling is not 1 or 2

Subsampling seems to work with 1 and 2, but gives incorrect
results for (3, 3) and (1, 3).

Fixes to argsort when axis is None.

Set stride to 0 on length-1 dimension in subtensor

If a CudaNdarray has the right shape to be used as storage for a
broadcastable Variable, its stride should be 0.

Closes gh-597.

C code cache version

Added tests to check that we free all intermediate results on the gpu.

This makes it easier to do some simplifications.

Like config.time_seq_optimizer, prints timing information on
EquilibriumOptimizer.

Also change a little bit how config.time_seq_optimizer prints some
informations.

Instructions for manual Win 64 installation

This is needed to be able to version the CURAND op and all other op that have a generic in its input.

This is needed to cache the c code.

Test fix

Make CLinker use preallocated output memory

Commented pycuda example that use the pycuda elemwise generator is it work only with old version of pycuda.