fix new crash in gpu optimiztion due to the new scheduler stuff.

Fix crash in gpu code in verbose mode when after using fprint_CudaNdarra...

We tried to copy to the CPU the dev_structure even when it was not allocated.

add gpu tags to gpu optdbs

Fix tests when cxx is available, but not blas

use make_(depends/dependence) rather than old fn

Np1.5

Schedule

Expand ~ in base_compiledir. clean-up after tests.

Change syntax for python 2.4

Make sure the output of perform has the exact right dtype

Conflicts:

	theano/tensor/extra_ops.py

Small improvements to memory management tutorial

fix grad for comparisons

Small

This should reduce pickling issues

It never schedules its own applys

This Linker never schedules ops on its own

replace code like

    order = fgraph.toposort()

with code like

    order = self.schedule(fgraph)

in Linker code

Use a relative tolerance in test_good

Add sort_apply_nodes function to gof/sched.py

it was causing compilation problem on windows for someone on the mailing list.