Profile node lru cache and revisit implementation

Currently the lru cache for nodes uses a vector-based queue to maintain the most recent items, which is updated each time a page is accessed.  However I would estimate there are typically ~1000 entries in each cache - which means each time a page is accessed it will copy O(4K) of data.

That is likely to be a significant cost - and it is all within a critical section.  A doubly linked list is likely to be much more efficient once N > 50 (fewer operations, but the queue is very cache friendly).

However we need some realistic test cases to be able to profile performance and test potential improvements.