I just come out a new algorithm to optimize traversing speed.

1. Sort every tree node's children by ts. (If we check PyTorch profiler and make sure the operators are already in time order, this can be omitted).
2. Sort all memory records by ts.
3. Traverse the tree only once, by memory records' time order. ("Walk" through the tree by time order leveraging node's ref to its father node, put the current memory record into its direct-father node; Then next memory record could start from this tree node, ...)
4. Traverse the tree bottom-to-up only once to summarize all columns.

Sorry that I think out late... Let's run the existing one firstly. If perf is not enough, let's optimize it later on.

you may need to add an ut to compare with autograd.profiler (refer to test_compare_with_aurograd.py)