disable gc.collect in the debug mode. This speed up tensor/tests/test_opt:test_fusion.test_elemwise_fusion from 30s to 3s!

make TensorDot and TensorDotGrad work with scalar axes. Add tests for that and test that is raise error.

added the theano.gof.destroyhandler.fast_inplace_check() fct that allow to pre-check is an op can be done inplace or not. It don't check all condition.

fix a bug in PatternSub when their is a constant in the pattern. In one case the optimization was not applied when it needed to be. Only affect the new optimization done this week.

use the new DebugMode machanism to disable error when an stability optimization remove nan for the local_{mul,div}_switch_sink optimization. Updated test to use it.

Implemented a new system that allow an optimization to specify a values_eq_approx to use for a variable.

This is usefull to make DebugMode accept that a stabilization optimization remove inf or nan from the graph.

in the stability optimization of the grad of log(erfc(x)) accept more equivalent type of graph and test that the opt are done in the stabilization phase.

add a STABILIZE mode and a stabilize optimizer. They allow to test that stabilization optimization happen in the stabilize phase. Not always the case now.

rename a variable and make the stability optimizer of the grad of log(erfc(x)) happen directly with the canonicalized graph

BUG FIX: the stabilization optimization of the grad of log(erfc(x)) was applied too frequently! That test for that bug