Do not use python code for elemwise with >= 32 inputs

NumPy is not able to deal with that many inputs to a ufunc,
and will segfault or raise an Exception.

[MRG] Add c_code for MaxAndArgmax

c.data + 1 return numpy.float64 when c.data is float32. But we compute i...

So when the test have floatX=float32, we have an error.

this remove a deprecation warning in NumPy as ndarray won't be
accepted as int anymore.

rebalence travis-ci tests.

[MRG]Fix opt profiler

Local subtensor of alloc

fix compilation crash recently introduced on the cuda code.

fix gh-1381 c linker crash with not used inputs.

Document how to add scalar op and random number distribution.

Fix typos, some rephrasing.

make EquilibriumOptimizer use a dict to map ops to their nodes rather than running everything on everything.

list() the iterator for local_optimizers