[sray014]

Adds WPCQPFReaper, which fetches NOAA WPC 2.5 km CONUS 6-hour QPF forecasts (GRIB2) from https://ftp-wpc.ncep.noaa.gov/2p5km_qpf/.

How it works

• Reads the server's directory listing to find the files for the requested issuance. init_time="latest" picks the newest issuance that has every requested forecast hour.
• Checks that init_time is 00, 06, 12 or 18 UTC and that forecast_hours are multiples of 6 within that cycle's limit (168 h for 00Z/12Z, 174 h for 06Z/18Z).
• Raises DataNotFoundError if any requested hour is missing, so summed tp can't silently come out too low.
• Downloads the files, converts longitudes to −180–180, applies any transformations to each file, then combines them along step.

[cheginit]

Cosecha mainly runs as scheduled AWS Lambda functions, where peak memory sets the configured size and the cost of every run. Please treat memory footprint as a design constraint for each reaper. Some rules of thumb:

1. Measure peak RSS. ds.nbytes undercounts. A full WPC issuance is 438 MB but peaks at ~1.25 GB. Check with resource.getrusage(resource.RUSAGE_SELF).ru_maxrss.
2. Reduce each file as it loads. Subset, select, and convert before combining. Subsetting after concat saves nothing.
3. Slices are views. A small isel/sel subset keeps its full parent array alive, so deep-copy it before holding on to it.
4. Don't hold inputs and output together. A list plus xr.concat peaks at ~2x the result, with or without a generator. Don't stack with xr.merge: it aligns every piece to the full index first (15x the peak of concat in a test).
5. Lazy-load heavy packages. import cosecha costs 263 MB against ~190 MB for what WPC needs, since pyiceberg (~54 MB), s3fs (~41 MB), and icechunk (~27 MB) load on every cold start. For a follow-up PR: a lazy public API like this template, and moving the icechunk/pyiceberg/pyarrow imports in reaping/base.py into the sow_to_* methods.

Requested changes (details inline):

1. Apply transformations per file (bbox peak drops from ~1.25 GB to ~550 MB).
2. Raise when explicitly requested hours are missing.
3. Bound forecast hours by the issuance's own cycle.
4. Keep the DataNotFoundError/APIError types.

Please also add a PR description and a README entry.