Improve priority queue use in C++ matching code

[unzvfu]

The use of the std::priority_queue in the C++ matching code (match_one_against_many_dice_k_top), is the main bottleneck when k is large. There a few deficiencies that may be negatively impacting performance. For example:

1. Extensive use of priority_queue::pop() in loops as though it were O(1) when it is in fact O(log n).
2. AoS memory layout of the queue elements (struct Node) when SoA would probably be better.
3. Possibly unnecessary maintenance of the queue invariant when k is small (a qsort at the end might do).

Some potentially useful resources:

• https://stackoverflow.com/a/10761286
• https://playfulprogramming.blogspot.com.au/2015/08/cache-optimizing-priority-queue.html
• https://github.com/rollbear/prio_queue

Aha! Link: https://csiro.aha.io/features/ANONLINK-70

[unzvfu]

Working on this in the fix-prio-queue branch; currently works but is slower for reasons that I don't currently understand.

[unzvfu]

Moving this to the icebox as my attempts to improve the implementation haven't yielded anything useful yet.