Audit OB1 + BTLs for multi-threaded out-of-sequence performance degredation

[jsquyres]

Per #2067, it seems like there might be optimizations (and/or some level of re-design) regarding how message / frag sequencing is done in ob1 to better handle multiple threads invoking MPI sends simultaneously (in an MPI_THREAD_MULTIPLE sense) to avoid the performance degradation associated with out-of-sequence scenarios.

Improving this will likely be a large effort. This issue is a placeholder for discussion and work in reducing the occurrences of OOS and/or the performance impact associated with OOS.