use single precision for most particle dsets

this should fix part of the I/O issues.
And, as a matter of principle, we don't really need double precision for
these quantities, for the sort of statistics we want.