Robustness guards for LSTDA

[jjokella]

Problem: NaN values to coupler after about 1 year of eCLM simulation with LSTDA.

Goal: Introduce more checks to the LSTDA-update

remark: co-authored by Claude-Code

Summary

1. spval filter on t_veg: Excludes non-vegetated patches (t_veg = 1e36) from
   TVEG averaging and updates. Falls back to TSKIN mean for fully non-vegetated
   gridcells instead of crashing.
2. Per-column update guard (col_updated): Prevents TSOIL increments from being
   applied multiple times to the same column when multiple patches share it.
3. NaN check on increment_factor: Fortran's abs(NaN) < threshold silently
   evaluates to .false., bypassing the cap. An explicit ieee_is_nan guard leaves
   the variable unchanged instead.
4. Temperature clamp [TKFRZ−130, TKFRZ+100]: Last-resort physical bounds check
   on all updated temperatures.