Not use emplace for editing vector elements

[csegarragonz]

Pretty simple, silly, yet nasty bug with the MPI implementation.

In two different std::vectors (which we previously filled by emplace_back-ing nullptrs, we used emplace(<offset>, <thing to be moved>) to modify the element at offset <offset>.

Turns out that this is not how emplace works. Citing cppreference:

Inserts a new element into the container directly before pos.

Thus, our vector elements were moved all around the place in a non-deterministic manner. The fact that this worked sometimes is because:

• We lazy-initialise these vectors. If the element is not initialised, we create a new one. Thus, every time we created a new element at the right index, and used it (just once). This only happened a limited number of times.
• Given that the rank distribution we test with is always lower-half host 1, upper-half host 2, and the vectors are sparse by nature, overwrites were occasional.

This would have worsen dramatically with more aggressive testing or different rank-host distributions. I checked, and this does not happen anywhere else in the codebase.

Should have read the reference better before hand.

[csegarragonz]

Caught this re-visiting the sources so figured out i'd just change it.

[Shillaker]

Good catch. With bugs like this it's good to add a test that would have failed before, but now passes (if possible). In this case perhaps it could combine calls to initRemoteMpiEndpoint and getUnackedMessageBuffer to check that things work?

[Shillaker]

Nice work on the tests. I've added some comments about how I don't think we need to mark them as test-only or throw exceptions, so once that stuff's been removed I think it's good to go.