Fix UW shallow convection mumin NaNs

[mathomp4]

This PR was from me asking GPT 5.6 Terra to look at the issues in #1466

It first saw an OpenMP issue that, while a bug, is not the cause of #1466. It then looked at compute_mumin2 and thought it could have odd behavior in some cases. So, I asked it for a compute_mumin3 and this is what it came up with.

I can say that ifx passed Debug with this.

---

This draft addresses the UW shallow convection mu solve that could produce NaNs in edge cases.

Why this change

• compute_mumin2 uses a Newton iteration starting from mulow.
• In some columns the iteration can move into a region where exp(-mu**2)/erfc(mu) becomes numerically unsafe, so the old solve can trigger NaN behavior instead of converging cleanly.
• That makes the limiter path fragile even though the intent is only to enforce the existing ufrc/cbmf constraints.

What changed

• Added compute_mumin3, a bounded bisection solve on [mulow, 3].
• Replaced the risky solve path with the bounded solver when the ufrclcl limit needs to be enforced.
• Kept the same physical constraint logic; this is a numerical robustness fix, not a change to the underlying shallow convection formulation.
• Also includes the OpenMP tracer-workspace fix in uwshcu.F90 so threaded execution uses private workspace arrays.

How mumin3 differs from mumin2

• mumin2: unbounded Newton iteration, faster when well-behaved, but it can overshoot into a numerically unsafe region and produce NaNs.
• mumin3: bounded bisection, slower but more conservative, and it stays inside a finite interval so it cannot chase the solve into the underflow/0-0 regime.
• In normal cases the two should give very similar mu values.
• The difference should mainly show up in pathological columns where the old solver was unstable or non-convergent.
• In those cases, mumin3 should return a finite bounded result instead of failing the limiter solve.

Validation note

• The main expected outcome is that NaN-producing cases now return a bounded, finite mu instead of failing the solve.

Acknowledgement

• This PR was developed with assistance from LLM-based tools, including GPT-5.6 Terra and GPT-5.4-mini.

[narnold1]

Can confirm that this fixes the "compute_mumin" issue, and has minimal effect on science output (based on BOMEX SCM case and a short AMIP exp).

[mathomp4]

As @narnold1 approves, I'll un-draft and @sdrabenh can test for his next update.