Fixed a rare logic error in the f_ground computation

When computing f_ground for cells or quadrants, the number of floating corners (nfloat)
is based on how many of the four corners have f_flotation > 0.  Arbitrarily, corners with
f_flotation = 0.0 are considered to be grounded rather than floating.

This creates a potential problem for a cell where 1 <= nfloat <= 3, but all the corners
that are nominally grounded have f_flotation = 0.  (This can happen in rare cases, because
of another piece of logic that caps f_flotation at values slightly greater than or less than 0,
so that small positive values can exactly cancel with small negative values.)
Then we have a log(0) in the integral formula, leading to f_ground = NaN.

The one-line fix is to treat any cell or quadrant with minval(f_flotation) = 0 as if nfloat = 4,
implying f_ground = 0.

This fix is answer-changing in a good way; it gives finite answers in these rare cases, instead of NaNs.

The fix was introduced partway through the ISMIP6 Antarctic production runs, so that these runs could finish.

libglissade/glissade_grounding_line.F90

@@ -971,12 +971,16 @@ subroutine compute_grounded_fraction(i,     j,     q,       rank, &
 
     ! Given nfloat, compute f_ground for each vertex
     ! First the easy cases...
+    ! Note: Cells with f_flotation = 0 are arbitrarily identified as grounded rather than floating.
+    !       This can be problematic if 1<= nfloat <= 3, but all the corners that are identified as grounded
+    !        have f_flotation = 0 (i.e., minval(f_flotation) = 0).  Then we should set f_ground = 0 (as if we had nfloat = 4).
 
     if (nfloat == 0) then
 
        f_ground = 1.0d0    ! fully grounded
 
-    elseif (nfloat == 4) then
+!!    elseif (nfloat == 4) then  ! need the extra logic in the line below
+    elseif (nfloat == 4 .or. minval(f_flotation) == 0.0d0) then
 
        f_ground = 0.0d0    ! fully floating