Opening Angle method optimizations and organization

[amoodie]

overview

Significant optimizations and organizations applied to the opening angle method. This is a breaking PR, because it fixes several bugs within the implementation. These bugs likely did not affect the results of any work relying on the computations, but were points where the code was inconsistent with intention (e.g., off-by-one errors).

changes

This PR also changes some default options to be consistent with the paper in some places, but optimized for speed and clarity in others.

• preprocessing is added to fill lakes (optional, enabled by default)
• the processing of water pixels and land pixels is now separated, but can still be accomplished with two calls to the dm.plan.shaw_opening_angle_method() with optional parameter query_set=lwi (land-water interface).
• the default "number of views" (numviews is changed to 1. This provides acceptable results in nearly all circumstances, with a significant advantage of speed. Support is maintained to use numviews=3, which is consistent with the original paper. Note, the old implementation was never using more than a single view; this PR fixes this, so that the opening angle is the sum of the numviews largest view angles.
• many variables have been renamed, to be self-consistent and consistent with the paper. Overall, I feel code readability is greatly improved.

speed

Speed also seems to be improved. In a simple example using the golf data and default options, the compute time is reduced from 0.79 to 0.28 seconds (excluding jit compile time!). The same test on this PR with numviews=3 is 0.43 seconds, due to an extra necessary call to np.sort.

PR also implements optional parallelization on the opening angle method calculation. Default is no parallelization, and an optional integer argument specifies the number of threads to use in computation. Parallelization here induces pretty low overhead for a smallish number of cores; therefore 2 cores -> ~2x speedup; 4 cores -> ~4x speedup.

I expect, but have not verified, that these benefits play out the same on larger data processing (e.g., timeseries with higher resolution data).