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

replacing the test id (module path and fct name) with the first line
of the docstring, making it hard to identify what version of the test
failed.

This is mostly for tests that are shared between tensor and cuda/gpuarray.