Document that `NaN`s are not suspicious

[nathanjmcdougall]

One principle for rastr development that I think we should document is that NaNs are treated as usual instead of unusual

Numpy tends to view NaNs as suspicious, and needing an escape hatch

Rastr tends to treat them as expected and first-class values (in a sense)

As a result, bugs like #395 are quite easy to creep in. It would be worth adding this information to the copilot instructions file, as well as CONTRIBUTING.md

Instructions file might be something like

Something like "always consider how NaNs are handled. We can expect that most rasters will have NaNs. Don't let them propogate disproportionately causing unexpected behaviour for operations and methods."