Crash weather-dl pipeline early for common class of errors.

Let's distinguish between at least two types of errors in the `weather-dl` pipeline: The first are pipeline execution errors. For example, the network could go down, or a shard of data could be corrupted during transfer. For these, we should retry a request with our current fetching logic. The second type of errors are user-produced errors. Issues like not setting up enough disk space, typing in a variable incorrectly s.t. EC doesn't know what data to give, etc. — these should fail the pipeline as soon as possible. Thus, we need to devise a way to not retry these types of errors and fail the pipeline, so we get an early feedback signal to fix it.