Tests have been rewritten, and they run so that
a CPU computation and a GPU computation are always
performed (separately) on the same input.

This allow to run the tests with Theano profiling
flags and then compare the execution time of
MaxAndArgmax (CPU) and GpuMaxAndArgmax (GPU).

Some code has also been modified in theano/tensor/basic.py,
related to MaxAndArgmax, to make the API more uniform
and to put most of axis checking in maxandargmax wrapper
instead of in make_node functions of (Gpu)MaxAndArgmax.

GpuMaxAndArgmax op is implemented in a separate file
in theano/gpuarray.

A test script has also been written.

To integrate this op in Theano, an optimization has been
added to ttheano/gpuarray/opt.py.

Added checks for taps values.

Fix remaining tests in debugmode

Update to fix python 3 error

Option to evaluate part of the max-and-argmax on cpu

Numpy imports

Fix h_softmax example data types

Fix grad of elemwise with boolean inputs.

Commented code removed, and assertion added to check that call to sys.getrefcount does nots changes the refcount.

Update test function test_pycuda_memory_to_theano() to fix Python 3 error.

The refcount of a variable seems to be different between Python 2 and Python 3. So it seems better to stock an initial refcount in a variable and then do the checkings with that variable instead of directly using comparing to constant values.

Tests passed with Python 3 (Anaconda):

```
$ nosetests --verbose --nocapture theano/misc/tests/test_pycuda_theano_simple.py
Run pycuda only example to test that pycuda works. ... ok
Simple example with pycuda function and Theano CudaNdarray object. ... ok
theano.misc.tests.test_pycuda_theano_simple.test_pycuda_memory_to_theano ... 4
gpuarray ref count before creating a CudaNdarray 4
strides (20, 5, 1)
gpuarray ref count after creating a CudaNdarray 5
gpuarray ref count after deleting the CudaNdarray 4
ok

----------------------------------------------------------------------
Ran 3 tests in 0.297s

OK
```

#2801:subtensor-incsubtensor

Import numpy

[REG, CRASH] fix crash related to float16 introduced in gh-5077

DebugMode is too strict here. It happens rarely, so I decided
to have a workaround for these specific tests.
Moving the definition of values_eq_approx outside of TensorType made it
possible.