Dask chunking in hf-sim

[lispandfound]

To support extremely large simulations with dense station grids we should introduce chunking using dask. The fastest way to do this would be:

1. Create an xarray dataset consisting of all the inputs (currently we use pandas dataframes, but these are not dask-ready unlike xarray datasets).
2. Chunk the inputs appropriately. Using auto chunking behaviour is not advised because the inputs are smaller than the outputs and so it will pick chunks that are too large.
3. Use xarray's map_blocks api, which allows us map over chunks in parallel. We would then iterate over each station in a for loop and dispatch to the fortran code.

While we are here I would make a further change to the hf sim code: write output to temp files in /dev/shm instead of /tmp. On some clusters /tmp is backed by a real filesystem, but we would rather write to memory (which /dev/shm is guaranteed to be).