Use more efficient algorithm to calculate dominators

Use Lengauer&Tarjan's algorithm to calculate the tree of
immediate dominators instead of the previous algorithm,
which gives us a significant performance improvement.

Also, instead of storing the dominators in a std::set we
now just store the immediate dominators and make lookups
in there instead. This is a significant performance improvement
as long as the dominators set are only iterated over or only
queried infrequently, as is currently the case, and also
saves a lot of memory in the case of large functions (previous
memory usage was quadratic with function size, now it is linear).

In cases where a lot of queries are made against the same set of dominators,
they can still be copied to a local std::set prior to that.

Author: hannes-steffenhagen-diffblue