Vert friction: Force FMAs in tridiag solvers

Switching the vertical friction loops from k/j/i to j/i/k replaced the
evaluation of `b1` by FMA with a simpler version, causing an answer
change when FMAs are enabled.

Although less efficient, this patch adds an always-false loop to trick
the compiler and force it to always execute `b1` by FMA.

Specifically, loops of the following form execute `b1` by FMA.
```
    do k=2,nz
      if (allocated(visc%Ray_v)) Ray = visc%Ray_v(i,J,k)

      c1(k) = dt * CS%a_v(i,J,K) * b1
      b_denom_1 = CS%h_v(i,J,k) + dt * (Ray + CS%a_v(i,J,K) * d1)
--->  b1 = 1.0 / (b_denom_1 + dt * CS%a_v(i,J,K+1))
      d1 = b_denom_1 * b1
      visc_rem_v(i,J,k) = (CS%h_v(i,J,k) + dt * CS%a_v(i,J,K) * visc_rem_v(i,J,k-1)) * b1
    enddo
```
Switching to j/i/k ordering allows the Intel compiler to cache `a_[uv](K)` for
use in the next iteration of `k` and evaluate `b1` by a single multiplication.

If we insert an impossible branch, such as the following:
```
    do k=2,nz
      if (allocated(visc%Ray_v)) Ray = visc%Ray_v(i,J,k)

      c1(k) = dt * CS%a_v(i,J,K) * b1
      b_denom_1 = CS%h_v(i,J,k) + dt * (Ray + CS%a_v(i,J,K) * d1)
      b1 = 1.0 / (b_denom_1 + dt * CS%a_v(i,J,K+1))
      d1 = b_denom_1 * b1
      visc_rem_v(i,J,k) = (CS%h_v(i,J,k) + dt * CS%a_v(i,J,K) * visc_rem_v(i,J,k-1)) * b1

--->  if (dt < 0) exit
    enddo
```
then it blocks the lookahead logic of the compiler and forces the FMA execution
as in the k/j/i version.

There is a moderate impact on performance.
```
Before:
                                     hits          tmin          tmax          tavg          tstd  tfrac grain pemin pemax
(Ocean vertical viscosity)             300      2.717543      3.805039      3.523935      0.174203  0.064    31     0   511
```
```
After:
                                     hits          tmin          tmax          tavg          tstd  tfrac grain pemin pemax
(Ocean vertical viscosity)             300      2.780148      3.999669      3.761651      0.210061  0.069    31     0   511
```
so this should only be considered a temporary fix until FMA answer changes are
permitted.

Author: marshallward

src/parameterizations/vertical/MOM_vert_friction.F90

@@ -753,6 +753,9 @@ subroutine vertvisc(u, v, h, forces, visc, dt, OBC, ADp, CDp, G, GV, US, CS, &
         ADp%du_dt_str(I,j,k) = (CS%h_u(I,j,k) * ADp%du_dt_str(I,j,k) &
             + dt * CS%a_u(I,j,K) * ADp%du_dt_str(I,j,k-1)) * b1
       endif
+
+      !### Force FMA evaluation of b1 by blocking lookahead with an impossible branch.
+      if (dt < 0) exit
     enddo
 
     if (associated(ADp%du_dt_str)) then
@@ -944,6 +947,9 @@ subroutine vertvisc(u, v, h, forces, visc, dt, OBC, ADp, CDp, G, GV, US, CS, &
         ADp%dv_dt_str(i,J,k) = (CS%h_v(i,J,k) * ADp%dv_dt_str(i,J,k) &
             + dt * CS%a_v(i,J,K) * ADp%dv_dt_str(i,J,k-1)) * b1
       endif
+
+      !### Force FMA evaluation of b1 by blocking lookahead with an impossible branch.
+      if (dt < 0) exit
     enddo
 
     if (associated(ADp%dv_dt_str)) then
@@ -1206,6 +1212,9 @@ subroutine vertvisc_remnant(visc, visc_rem_u, visc_rem_v, dt, G, GV, US, CS)
       b1 = 1.0 / (b_denom_1 + dt * CS%a_u(I,j,K+1))
       d1 = b_denom_1 * b1
       visc_rem_u(I,j,k) = (CS%h_u(I,j,k) + dt * CS%a_u(I,j,K) * visc_rem_u(I,j,k-1)) * b1
+
+      !### Force FMA evaluation of b1 by blocking lookahead with an impossible branch.
+      if (dt < 0) exit
     enddo
 
     do k=nz-1,1,-1
@@ -1232,6 +1241,9 @@ subroutine vertvisc_remnant(visc, visc_rem_u, visc_rem_v, dt, G, GV, US, CS)
       b1 = 1.0 / (b_denom_1 + dt * CS%a_v(i,J,K+1))
       d1 = b_denom_1 * b1
       visc_rem_v(i,J,k) = (CS%h_v(i,J,k) + dt * CS%a_v(i,J,K) * visc_rem_v(i,J,k-1)) * b1
+
+      !### Force FMA evaluation of b1 by blocking lookahead with an impossible branch.
+      if (dt < 0) exit
     enddo
 
     do k=nz-1,1,-1