[FEATURE_REQUEST] Implement Faster OpenSFM by Yan Noun

[ichsan2895]

This PR makes OpenSFM significantly faster in mapper phase.
I have test it by myself.

Dataset 1st = personal dataset with 17 images
Before

After (almost 1.6x times speed gain)

Somehow, it does not merge yet into mapillary/opensfm with unknown reason.

Another idea from me that worked well:

If you compile latest commit of Ceres with CuDSS (Cuda Direct Sparse Solver) library, it supports GPU acceleration in mapper.
Just change the code:

Before (southbuilding from colmap example dataset that contains 128 images)

  options.sparse_linear_algebra_library_type = ceres::SUITE_SPARSE;

After

  options.sparse_linear_algebra_library_type = ceres::CUDA_SPARSE;
  options.dense_linear_algebra_library_type = ceres::CUDA;

I hope it merges into ODM