raise the number of allowed optimization application before raising an Equilibrium max'ed out. This remove useless output in test.

re-enable the grad on max in some case. The new max_and_argmax optimization do the manually done optimization that disabled the grad in some case.

make CAReduce raise the same error type as numpy when the axis is wrong and make it check negative axis.

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