Work out how to scale `.verify()` to lots of inits

Right now .verify() does not scale well to many inits. Currently I am trying to verify an S2S system (that was setup like SubX) with ~1,000 inits and 45 leads. I am just working with a time series and it’s prohibitively slow to even set up the dask graph or to just do in-memory. I think this has to do with constructing a huge dictionary that has the right dates at every init to align.

I think we need to evaluate the shift in datetimes while evaluating to make dask happier. Although I think I went with this dictionary method since you’d only have to do the shifting once, and to figure out which inits to keep for the various alignment strategies.

[ahuang11]

I think you need to do %lprun -f he.verify he.verify(**kw).compute() for lprun to work

[aaronspring]

.isin() is the bottleneck here. Maybe dict helps, see https://twitter.com/jesperdramsch/status/1486235033153282049?s=21