Add full support for np.memmap in RasterData

[leandroleal]

Description

Currently, the RasterData attribute array refers to regular np.array. Aiming to speed-up all data array operations we need to use np.memmap arrays, taking advantage of pre-allocation memory spaces and concurrent writing in a single shared data structure.

This re-factoring must include:

• skmap.io.read_rasters
• skmap.io.RasterData.read
• All skmap.io.process (Calc, TimeAggregate, etc)

[leandroleal]

Consider switching the group axis to the first array dimension, taking advantage of memmap.resize.