Keep only best SIFT features and other fixes

[pierotofy]

• In feature extraction using SIFT, keeps only up to 1.5x of the
  best min-num-features features (matching the behavior of HAHOG).
  Very large images tend to result in VERY large number of features, which
  slow down matching without adding any benefit. So this will speed up
  the SFM step while using SIFT.
• Makes the --matcher-order parameter kick-in only if the dataset is
  not georeferenced. There's really no reason to apply matcher-order to
  georeferenced datasets, where graph or distance matching works better.
• Makes the start-dev-env.sh script work on MacOS.