Speed up the i/o of the drainage network

[AlexLipp]

Create a load_from_node_arrays function in autocatchments so the network stack doesn't need to be built every run.

[AlexLipp]

Such a function is now implemented in autocatchments and results in a 2x speed but result is undesirable as:

• The plaintext ordered node arrays take up much more harddrive than present .nc file.
• Its not transparent how they are calculated but the .nc is a widely interoperable file.

Instead, more desirable speed ups can be gained by:

• Implementing the conversion from D8 to receiver node in a lower-level language (cython) (listed here)
• Not using the slow RasterModelGrid class but developing a specialist one with lower overheads (listed here).

[AlexLipp]

PR #8 has partially increased speed up (3-4x) by use of pickling but it can be improved more by using a lower overheads object than the RMG

[AlexLipp]

Easy speed-ups could be gained by:

1. Downsampling the input DEM and preprocessing it in advance (i.e. sink-filling)
2. Cutting the study area down to be more tightly along the Thames Basin

More challenging speed-ups could be gained by cythonizing the channel-profiler