StoredRange can store an optionally-deleted unique_ptr

[jwpeterson]

• Also add some libmesh_asserts in StoredRange::reset() to catch the case
  where a range is reset() but the _objs pointer was not initialized.
• Use lambdas to define the custom deleter. The decision to delete/not
  is made at construction time, and does not change during the object's
  lifetime.
• One thing I am not 100% happy with is that we still need to call 'new'
  when allocating the class unique_ptr, since one cannot really use
  std::make_unique when a custom deleter is employed.
  https://stackoverflow.com/questions/21788066/using-stdmake-unique-with-a-custom-deleter
• Use ptr_type typedef for improved readability
• Add required C++ includes.

[moosebuild]

Job Coverage on 97fafb9 wanted to post the following:

Coverage

	efa6ff	#3332 97fafb
	Total	Total	+/-	New
Rate	55.73%	55.72%	-0.00%	100.00%
Hits	44901	44899	-2	5
Misses	35674	35674	-	0

Diff coverage report

Full coverage report

This comment will be updated on new commits.

[jwpeterson]

So we have a randomly failing Min clang target, is that correct?

[roystgnr]

Yup. In MetaPhysicL. I'll see if I can replicate it locally.

[roystgnr]

Well, that's fun. With clang-11, I can run MetaPhysicL test/instantiations_unit in a loop and get a SEGFPE within seconds, after 60 or 70 successful runs. This is supposed to be deterministic code. Looks like I can get the same with clang 12, so just seeing it from the Min clang target may have been luck. Let me see what other compilers work vs fail.

[roystgnr]

I'm getting the same non-deterministic failures from clang 13 and 14, and gcc 9 and 11; this has got to be our problem, not a compiler issue.

Unless it's an issue with something else in the environment? I went to an older Linux box to try clang 10, the compiler in our failing CI runs, and that turns out to be happy to make thousands of runs without a hiccup.

[roystgnr]

... instantiations_unit is outputting uninitialized values. That's undefined behavior, and it's non-deterministic in the heap-allocated cases because the heap might be pulling memory we've already used for something else, and some of that memory might be effectively getting a reinterpret_cast to NaN.

I'll fix it.

[jwpeterson]

OK, sounds good. From what I could see, the metaphysicl test suite doesn't yet output the name of the test which failed (?) but that would be useful for trying to diagnose problems via CIVET... since I don't think the unit test log can be downloaded as part of the results.