Our use of filters in CTSM degrades performance and doesn't allow any loops that use them to vectorize

A core aspect of CTSM's architecture is the use of filters for looping. These were originally put in place to help performance - and particularly (I think) to help vectorizability of the code. I've wondered, though, if they still help performance overall on current architectures, compared with a simple loop with a conditional inside (i.e., we'd still only operate on the same points, but we'd do so via a conditional rather than via indirect indexing). I'm wondering if the indirect indexing might hurt more than it helps.

I could especially imagine filters hurting more than helping in cases where the subgrid structure is set up to mainly contain points of a single column type (e.g., all natural vegetation), like what @barlage is doing for comparison with Noah-MP.

It could be worth timing a few representative samples with the current filter-based approach vs. a simple loop with nested conditional.