Make obs perturbation grid-cell lookup in get_obs_pert() layout independent

[amfox37]

Summary

In get_obs_pert(), each obs was mapped to a perturbation grid cell with get_ij_ind_from_latlon(pert_grid_lH, ...). The local grid pert_grid_lH is built per processor from the local obs extent, and its ll_lon/ll_lat are computed in single precision (e.g. ll_lat = pert_grid_f%ll_lat + real(ind_min-1)*dlat) and are subject to roundoff. An obs lying very close to a pert grid cell edge could therefore be mapped to different cells, and so receive different perturbations, depending on which processor handles it. As a result, the analysis depended on the EnKF layout (number of processors and AnaLoadBal partition). Within a single run, the same obs could also get different perturbations for tiles analyzed on different processors.

This PR finds the cell on the global pert_grid_f, which is identical on all processors, and converts it to the pert_grid_lH index with the same integer offsets (i_offg/j_offg) that propagate_pert() uses to extract the local perturbation window. The perturbation fields themselves were already layout independent; only the lookup was affected.

Evidence

CF0360, 24 members, H SAF ASCAT, 240 ranks, comparing the default count-based AnaLoadBal partition against an experimental cost-weighted partition, with the same executable otherwise:

• Before: obs and forecasts were identical between the two partitions, but the analysis differed at 45 obs / 368 tiles in one cluster (Venezuela). A diagnostic print traced this to a single METB obs at lat 5.1666660, which is 3.8e-6 cells from a 1/6-degree pert grid edge. It mapped to pert-grid row 571 on some processors and row 572 on others, where the local ll_lat rounded differently, so it got different perturbations.
• After: the two partitions are bit-for-bit identical: ObsFcstAna, catch_progn_incr and all restarts at 03z and 06z.

Impact

• Not zero-diff: results change slightly wherever an obs lies within roundoff of a pert grid cell edge.
• This removes the obs-perturbation source of layout dependence in the analysis. It does not make the whole model layout-independent: with an Aggressive build, runs on different processor counts still differ slightly for other reasons.