Multi-threaded BLAS

[SeguinBe]

By default, the request to add an image in the database needs to:

• Download the image
• Extract the CNN signatures
• Input them in the DataManager

Currently, most of the time (several seconds for a single image) is spent for the CNN computation if BLAS is not multi-threaded. Linking numpy with OpenBLAS should do the trick efficiently.

[SeguinBe]

Features computation is now deported to a task-queue on a separate machine. So efficiency is not as critical.

Still haven't managed yet to use OpenBLAS or MKL on Mac