Remove the linear scanning in the memory tracker performance

[qining]

When a new dirty page is to be added to the memory tracker, the tracker now scan through all the recorded dirty pages first, which spends O(n) time. This can be improved by implementing a simplified hash map (linked list + static sized array).