Use the new pycolmap.Sift interface

[sarlinpe]

colmap/pycolmap#57 introduces a unified interface for CPU SIFT (vlfeat) and GPU SIFT. The later is 20x faster (30 images/s instead of 1.5) but seems slightly less accurate. Test with pipeline_SfM with netvlad, sift, NN-ratio:

previously
	num_points3D = 32407
	num_observations = 186116
	mean_track_length = 5.74308
	mean_observations_per_image = 1454.03
	mean_reprojection_error = 0.871393

new CPU
	num_points3D = 32388
	num_observations = 186084
	mean_track_length = 5.74546
	mean_observations_per_image = 1453.78
	mean_reprojection_error = 0.872367

new GPU
	num_points3D = 31489
	num_observations = 178612
	mean_track_length = 5.6722
	mean_observations_per_image = 1395.41
	mean_reprojection_error = 0.870559