Investigate AWS lambda execution

Brainstorming.

If #3 works well for windowed reads, create a lambda function for `Counts` and deploy to a variety of memory configurations (ie, 0.5GB, 1GB, 1.5GB). Route requests through a load balancer which can take the input geometry and the target raster metadata (cell size + data type) and evaluate the amount of memory needed to read in the bounding box window - then route to that configured function for processing.

Pros:
- Scalable for concurrent requests, no server to overwhelm
- Minimizes cost per request
- Could scale to many generic functions, ie `Counts`, `MapAlgebra_Add`, etc

Cons:
- S3 networking latency
- Can GDAL and all the requirements actually fit?
- Cold start up time adds to request time
- Lambda memory seems to max out at 1.5 GB, not sure how limiting that would be for certain queries given other system memory requirements