Chunked fields operational

[delandmeterp]

This PR splits the field_data into chunks.

Field.data is not a np.array anymore but a lazily loaded da.array.
If a JIT kernel needs to interpolate data within a chunk, this chunk is loaded into memory.
In Scipy, pset.execute() first converts all Field.data into np.arrays.

To do

• What to do with Field.gradient()
• Cut Field.chunk_data() into an initial call and what is called during simu
• Cleaning
• Field.chunk_data should be moved to Field.grid.chunk_data
• We should catch the following situation:

pset.execute(kernel=InterpOnlyU, ...)
pset.execute(kernel=InterpOnlyUandV, ...)

where grid.load_chunk_data will not be matching with loaded V chunks
-> solution here is to attach load_chunk_data to the field itself...
NO, it does not work, since really not efficient
-> we keep it on grid, but in chunk_data, we also load the data also if it's not loaded yet, but then we also need to set chunk_data back to None at the pset.execute() if it't not used by kernels!

• What to do with data_array interpolated in scipy. it can happen, especially when called out of a pset.execute (at particle.variable init for example). Putting tests for da.array and compute everywhere?
• repeat kernel are correctly interpreted with pset.execute(dt=0) ?
• add 3D example plot to documentation notebook