[ENH] avoid "load from repo" with stallion data, replace in testing

[fkiraly]

The get_stallion utility downloads the stallion data from the repository, each time it is called. This seems wasteful of download traffic, and ceases to work without an internet connection.

At least in testing, this should be replaced - perhaps by a synthetic dummy dataset of comparable format.

Further, for the user-facing case, we should consider moving it into the package - it is 1MB and 0.8MB after compression, so not too large - but also not minimal size.