Add TensorConstant check in check_node_state method.

refactor test and print the line to check with the expected answer.

increase the complexity of test.

1. remove compute_map.copy from vm.py
2. add new compute_map in compile.py
3. make code more simple

1. remove global variables
2. modify discription
3. use new way to track order

1. compute_map reset process in min_memory_generator

1. use cpmputed_map as a argument for min_memory_generator.
however which it will works remains doubt, the question I will email Fred, and Arnaud

1. Simulate execution step in generator.
2. Use new memory usage counting method

thx to Fred

1. make global variables correctly
2. use fgraph.apply_nodes replace the previous fgraph.nodes

the error is mainly caused by the check node is not work well

in previous version the memlist will store the current_mem if it's 0, but we don't need that value, so I just remove it.

convert set into list for node_list

1. move count_minimum_peak in profile_memory
2. use nodes_mem in fct

1. remove the best order, cuz we won't use it now.
2. move the count_minimum_peak to the right place

1. Call the new method in summary_memory
2. Add new print line for minimum peak
3. Use fgraph.nodes as inputs

2. create compute_map and node_mem in new method

Fix tests in buildbot and memory leak with allow_gc=False