Previously, if l0 was None, we tried to calculate a default initial level. This was
problematic if yobs had no valid points from which to calculate a mean. Now, if yobs
has no valid points, default to l0=0. This is a bit arbitrary, but it is necessary
to prevent a NaN-valued l from being passed back out. Other defaults are equally
arbitrary.

Also, I noticed that the `s` variable was getting converted from a NumPy array to a
list just before the method returned. For consistency with other outputs, we no
longer convert, and just return a NumPy array.