Created by: liu-plus-wei
- remove the MemoryBlock Desc logic, only record the the mapping of raw pointer to MemoryBlock for the allocated block
- remove the UpdateGuard and CheckGuard logic
- remove buddy_allocator_test_data and related code
For GPU: 5.18s => 3.42s (without absl)
For CPU(no regression): 3.438s => 3.420s (withou absl)
When counting the the original time cost (5.18s, 3.438s), I have disabled UpdateGuard/CheckGuard which will cost another 2+ seconds.