`glbvertex` might need zero check

[bennn]

I was testing Finch by randomly injecting NaNs and would up with a bounds error in grid.jl (see utahplt/FloatTrackerExamples#5).

The bad bound was 0 and it came from grid.glbvertex. (Caused somehow by a NaN on l.248 of grid.jl.)

That matrix starts off with all zero entries [link]. It should get initialized later on [link]. Perhaps it (or loc2glb) could use a check to make sure the initialization happens.