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.

PG-PuReMD: switch from using strtok_r to strtok to avoid reliance on GNU extensions with building for the C11/C++11 standard.

PG-PuReMD: only copy host generated grid to device on reneighboring steps (after rebinning atoms and regenerating the grid). Fix small NVE integration bug (reallocation checks before copying new atom info to device).

PG-PuReMD: change CUDA kernel error checking to only make calls to cudaDeviceSynchronize when in debugging mode (non-debug builds thus have better performance with kernel asynchronous launch behavior in single default stream). Change file I/O to only flush when in debugging mode (non-debug builds allow the I/O operations to be grouped for better performance).

PG-PuReMD: add support for symmetric, half stored format (SYM_HALF_MATRIX) of the sparse matrix for the charge model (add initialization routines and fix-up solver).

PG-PuReMD: rework custom reduction functions to use less shared memory and to have the correct thread and block counts. Fix bug in charge matrix initialization (under-allocated space previously). Split charge matrix and bonds/hydrogen bonds memory management routines to mirror the similar splitting of the initialization routines. Code clean-up related to half vs. full list and sparse matrix formats.

PG-PuReMD: improve performance by conditionally re-running initialization kernels within Cuda_Init_Forces based on out-of-memory conditions (for charge matrix, bonds list, hydrogen bonds list).

Tools: add PuReMD custom geometry replication to geo_tool.py. Remove older awk scripts. Fix issue with silica 6000 atom PDB file (CRYST1 lines not 70 characters).

Tools: update run_sim.py to generate more robust Slurm and Torque/PBS job scripts with the MPI+CUDA code (added extra option for specifying additional flags to be used with the command used to invoking the mpi+x code).

PG-PuReMD: fix performance logging code around GPU code. Rearrange header files to allow some preprocessor definitions to be defined via options in configure script. Rework Autoconf and Automake code to allow passing flags directly to nvcc compiler wrapper (for CUDA code). Enable C11 and C++11 standard targets during compilation.

PG-PuReMD: fix compilation error (remove variable). Fix issue with dual CG solver for QEq (local arithmetic in SpMV was incorrect). Revert CG solver convergence criterion to use norm of the preconditioned residual vector.

PG-PuReMD: small fix for initializtion routine performance logging (timers not reset after outputting to log file).

PG-PuReMD: corrections for utilizing a SAI preconditioner in MPI code (errors introduced during previous code merges).

PG-PuReMD: revert CG solver convergance criterion to used preconditioned residual norm. Re-enable dual charge solver for QEq.

PG-PuReMD: rework performance logging code in linear solver routines to avoid excessive MPI communications.

PG-PuReMD: adjust performance logging to compute mean timings across all processors. Other code clean-up.

PG-PuReMD: fix issue with certain data structures not being reallocated when the local number of atoms owned by a processor increases after exchanging messages with neighbor processors. Fix an issue where the MPI send message buffer may be overwritten by a received message (Coll). Small fix to not overwrite the program status return value with local function return values. Other code clean-up.

Tools: fix geometry file extension auto-detection for run_md_custom. Add missing restart-related control file keywords.

PG-PuReMD: fix out-of-bounds mmeory accesses and unitialized data usage in energy and force tabulation routines for van der Waals and Coulomb interactions.

PG-PuReMD: fix issue with some interaction lists not being initialized on first step of simulations from restarted runs. Fix typos in C++ code for utilizing C code (__cplusplus). Clean up C code for utilizing C++ code (unconditional extern's for unmangled member naming). Ensure that CUDA thread and block sizes are correctly set. Tweaks to memory transfers and allocation logic in integration code (GPU). Fix logic error with charge matrix allocation in GPU code. Other code clean-up.

Tools: fix geo_format for run_md_custom (geo file is mandatory, so the file type extension should always happen).

PG-PuReMD: add BiCGStab solver for GPU code. Corrections to CUDA block and thread sizes for some kernels. Fix some variables being used with uninitialized values.