More stable reduce operations by default

Fix check that was too restrictive

Update installation instructions for Anaconda

Checks the dimensions and bcast in inc_subtensor

Make sure min <= max in test_clip

Do not create a complex tensor with zeros, then cast it into floatX,
create the floatX zeros directly

If acc_dtype == output.dtype, nothing changes.

This dtype is to be used as accumulator, instead of self.dtype as was
previously the case. Then, the result is cast to self.dtype if needed.

Also implement the perform version using it.

Test default blas flags before using them

before returning it as default value

Fix align

Issue #783: python3 compatible - remove relative imports.

Lazy config default value

Also add installation/config script for Windows/Anaconda

Fix a test when g++ is not available.

Add triangle/nonzero functions

Try to fix the problem that depending of the inputs strides, the outputs strides of dimensions of size 1 get strides that make it considered not aligned by the PyArray_ALIGNED macro.