Performance of EAKF when the sd is 0

When using the QCEFF with BNRH there is an early exit from the sequential_obs loop when the probit transform fails:

https://github.com/NCAR/DART/blob/ba604a564e5950d74b78c6b057da13ec5ebc01ee/assimilation_code/modules/assimilation/assim_tools_mod.f90#L705-L715

to_probit_bounded_normal_rh fails it sd <= 0.0_r8 (so -99 failed, or 0). For observations such as radar_clearair_refractivity, there are lots of forward operators where the result is 0 across the ensemble. When this happens the fwd ops fail to_probit_bounded_normal_rh

Bailing out of sequential_obs at this stage saves a lot of unnecessary get_close calls and therefore a lot of runtime. I believe the get_close calls are unnecessary because if there is no spread,
https://github.com/NCAR/DART/blob/ba604a564e5950d74b78c6b057da13ec5ebc01ee/assimilation_code/modules/assimilation/assim_tools_mod.f90#L1497-L1501
the assimilation increments are zero
https://github.com/NCAR/DART/blob/ba604a564e5950d74b78c6b057da13ec5ebc01ee/assimilation_code/modules/assimilation/assim_tools_mod.f90#L1548

so no need to do any get_close to find state to not update.

Here is a profiling pic where about half the obs are 0 spread. Eakf is top, same shape, but you can see sequential obs do (the green bit) is stretched.

Note this effect is amplified by the get_close_initialization #910 cost, but even if you remove that there is still a considerable performance penalty for eakf vs bnrh when there are bunch of fwd ops with sd = 0.