Would an LRUCache add anything to `convolve_2d`?

[phargogh]

Caching is an important way that we can avoid disk I/O, as we have been doing in routing for some time now. GDAL states that it already has an LRU caching mechanism (and has for a long time), and also python now has an LRUCache in the standard library.

So, do we gain anything performance-wise to implement a python block cache of numpy arrays?