We should use optimization tags to enable, disable them. This also could cause problem on test in case there is a bug, teardown would not be executed per James comments.

-renamed TestGer_local_gemm_to_ger to TestGer
-changed the Ger.infer_shape methode with ger.perform
 This make test of cpu and gpu ger op easier and follow what is done most of the time
-Moved tests in TestGer_make_thunk to TestGer and use theano.function
 This was needed to make them work on cpu and gpu
-make TestGer always call call the compiled function

This make this test executed on the gpu too.