Minor preprocessing bugfixes

Running through the docs in a fresh environment I noticed a few very small bugs.

Config:

1. Rename split -> n_split for Zprime sample to correctly split.
2. Add merge_output for testing samples. Merge steps in the documentation crash otherwise. If the merged test files are never needed, perhaps I can just modify the documentation instead to remove the merge step for the test samples.

Preprocessing:

3. Remove if statement, so that if n_split = 1 in the config, the input file is still globbed correctly. (I ran into this bug before fixing 1. (without renaming to n_split this defaults to 1, and subsequently no input files were found, as -file_0.h5 is appended to the file even when n_split = 1).
4. Ensure scale dict dir exists before attempting to write.

Tagging @dguest @mguth, thanks.