Use vertices to define low_order_key() for neighbor-finding and mesh stitching

[roystgnr]

This resolves #3738 for me

[roystgnr]

Looks like the new tests are passing in serial but still failing in parallel, even on a ReplicatedMesh. That's unexpected.

[roystgnr]

Our _boundary_side_id is sorted by pointer, so it differs from processor to processor, but in BoundaryInfo::clear_stitched_boundary_side_ids with clear_nodeset_data we just operate on the first element from which we see a side ... so if that's the higher-order element then we clear the higher-order side nodes' ids and if it's not we don't. And it looks like we're clearing all node boundary ids there, even if there are more than just the ones we used to stich. Should be easy enough to fix both problems at once.

[roystgnr]

The Test MOOSE min clang failures are an unrelated problem downstream. If everything else seems happy here then we're ready to merge.

[moosebuild]

Job Coverage on 43c742d wanted to post the following:

Coverage

	a38ab5	#3752 43c742
	Total	Total	+/-	New
Rate	62.43%	62.29%	-0.14%	84.09%
Hits	68374	68267	-107	74
Misses	41145	41334	+189	14

Diff coverage report

Full coverage report

Warnings

• New new line coverage rate 84.09% is less than the suggested 90.0%

This comment will be updated on new commits.