Implemented grad for cudnn softmax.

Added test to ensure SoftmaxGrad to DnnSoftmaxGrad is not applied when cudnn is excluded from optimizations.

Added optimization that converts SoftmaxGrad to DnnSoftmaxGrad and associated test to make sure it is applied correctly.

allow cxx flag to be full path to compiler

Conflicts:
	theano/gof/cmodule.py

Dnn default and doc

Prevent computations in float16 in scalar and elemwise

Catch exception if open() fails as well.

Also:
- add tests for "inv" op
- do not test inplace if the input is int and the output is float
- remove a couple of redundant dicts

Otherwise, the expected values could be computed in float16,
which would not be precise enough, and which would not be supported
as a dtype by Theano.

Since many elemwise ops do not call the scalar op's impl() method,
we needed to also change the perform method when numpy built-in
ufuncs are used.

Ops that are defined with upcast_to_float should upcast to
float32 or float64 minimally.

Add tests for the cases where the inputs were int8, which is when
float16 values appeared.

Drop support for python 2.4.