Daymet (and other dataset) subsetting performance

[stuckyb]

Daymet dataset subsetting for monthly data is currently implemented such that the entire dataset for a year is clipped, the result is cached, and then individual months are returned from the cached result. This does not appear to working as expected: requests for clipped monthly data either run indefinitely or crash.

[stuckyb]

After some investigation, the performance problems appear to be caused by rioxarray attempting to load the entire dataset into memory before subsetting. The from_disk parameter should address this, but it appears not to work for NetCDF files. Turning that parameter on and switching to the TIFF versions of the data files fixes the problem. This is implemented in c05b49a. I am not closing this issue, though, because the question of optimal strategy still requires further investigation.