Side input datasets must be materialized or else memory leaks

[langmore]

There is a memory leak users will cause if side inputs like land_sea_mask or thresholds are not materialized.

Here is the step-by-step for land_sea_mask.

1. inside ComputeStatisticsAggregateAndPrepareForCombine.process, the line aggregation_state = self.aggregator.aggregate_stat_var(stat) makes a non-materialized aggregation_state via
   1. binning_method.create_bin_mask(stat) which creates a non-materialized mask
   2. xr.dot(stat, *weights, *bin_masks, dims=reduce_dims_set)
2. During CombiningSum, the accumulator is not materialized
   1. Memory grows > O(N) (because a growing Dask graph is being built?)

I see some solutions

1. Document and logging.warning to encourage users to materialize
2. Materialize side inputs (land_sea_mask, thresholds) when they are passed to the user class in __init__
3. Materialize as late as possible, in case the side input is sliced first.