Try to improve update_last_irreversible_block performance

[pmconrad]

Attempt to resolve #1105

Speed increase is insignificant (4 seconds for 22M blocks).

The first commit compares the new calculation method against the old, the second commit removes the old method.

[abitmore]

I think this is the main place where the performance got (slightly) lost, because we walk number_of_active_witnesses/2 steps on each new block. Each step costs log(number_of_all_witnesses).

My thoughts:

• maintain a small set containing the witnesses that produced a block after LIB and are still in the active witness list;
• maintain a pointer/iterator to the current witness that holds the LIB, and probably another pointer/iterator to its next witness in the set for better performance;
• when a new block comes,
  • if its witness is already in the set, LIB does not change;
  • otherwise, add the new witness into the set,
    • if the new set size is smaller than or equal to number_of_active_witnesses/2, LIB does not change;
    • otherwise, move the pointer/iterator and LIB to the next witness in the set, and remove the old one from the set.

With this approach we walk one, two or three steps on each new block within a smaller set with cost log(number_of_active_witnesses) (one for the lookup (if a witness is already in the set), one for adding it, another for removing the old). This can probably be done with a secondary index to keep data consistency. Not sure how much performance gain it will bring though.