nvtx+summary: improving nvtx by tracing initialization, which also fixes potential bogus first measurement of Projection for the summary

To enable it use undisclosed CMake option -DUSE_NVTX=ON (by default it is OFF)

REVERT: temporarily disabled fastest CUDA detection, since it strangely causes error on dvl machine (reverted from commit 1459580e)

adding older preliminary data for the initial Autotuning algorithm. Results obtained on dvl01 machine