PG-PuReMD: fix warnings (and errors) from several uninitialized memory regions due to multiple stream change. Fix race condition with valence angle computation due to mistake with dynamic shared memory. Add control file parameter for setting the number of CUDA streams. Other code clean-up.

PG-PuReMD: fix issue with post-processing step with QEq pseudo-charges (too little storage in q array). Fix issue with device-wide CUB reductions and prefix sums reallocating device memory too frequently. Rework CUDA-aware MPI code to fix issues (still seeing out-of-memory errors on Ascent cluster at OLCF).

PG-PuReMD: ensure MPI pack/unpack CUDA kernels are run in the correct stream. Other changes to bring the CUDA-aware MPI functions for the charge solver in line with their MPI counterparts.

PG-PuReMD: utlize multiple CUDA streams to run independent kernels in the ReaxFF+QEq task dependency graph.

PG-PuReMD: fix stream issue with cudaMemset's happening in default stream (switch to cudaMemsetAsync). Rework memory management in light of this change.

WIP: finish the explicit single stream port by using cudaMemcpyAsync and cudaStreamSynchronize calls.

WIP: changes to support multiple CUDA streams. This commit contains porting of all kernels from the default stream to a single non-default stream. This commit has bugs regarding cudaMemcpy's still being in the default stream (and synchronous).

PG-PuReMD: implement MPI+CUDA versions of SDM, PIPECG, PIPECR for charge solver methods (along with missing dual QEq versions).

PG-PuReMD: use a warp of threads for addtional bonded interaction computations (bond orders, bonds, lone pair/over/under).

PG-PuReMD: use a warp of threads for apply bond derivatives to forces and simplify these code paths.

PG-PuReMD: fix incorrect summations in kinetic energy and temperature calculation functions. Replace several hand-coded device-wide reduction kernels with CUB libary reduction.

PG-PuReMD: split bond and hbond initialization kernels for future optimizations. Cleanup CUDA code related to far neighbor list being in full format (remove unnecessary conditionals and arithmetic).

sPuReMD: make zlib dependency optional (trajectory file compression). PG-PuReMD: remove erroneous zlib header file inclusion.

PG-PuReMD: handle a few missed reductions in nonbonded interactions more efficiently. Clean-up bonded interactions functions and do not export kernels via header files. Increase C++ standard to c++14 for CUB (Thrust dependency). Continue with bundled CUB version and transition to version included with CUDA SDK in the future.

PG-PuReMD: use atomics for both partial energy and force accumulation (small reworks to reduce number of atomic operations). Remove several unneccesary cudaMemset calls. Other code clean-up.

Test: fix Python wrapper for QM/MM (interfaces out-of-date). Fix Fortran tester bugs. Add test artifacts.

PG-PuReMD: fix BGF parser for >1 process (correct atom counts within local simulation box boundaries).

PG-PuReMD: fixes to NVT with Berendsen thermostat. Use updated constants from sPuReMD. Add BGF geometry file parser. Other code clean-up.

PG-PuReMD: in-progress implementation of CUDA-aware MPI for charge solver communications using OpenMPI (default off).

PG-PuReMD: fix van der Waals and Coulomb force computations for 1 warp per atom implementation, and enable this implementation by default.

PG-PuReMD: use registers for local arithmetic in CUDA nonbonded calculations and fix shared memory usage in 1 warp per atom implementation.

PG-PuReMD: switch timers to use CUDA events and add additional timers for kernels to match MPI code. Other code cleanup.

This reverts commit 6a1385e9.

PG-PuReMD: align CSR sparse matrix data structure to 128-bit/256-bit segments for memory coalescing (for 1 warp per row SpMV implementation). Change default interaction list and sparse matrix allocations to be multiples of 32 for aforementioned memory coalescing.