ENH: Promote `MRFImageFilter::SetClassifier` to virtual and `m_ClassifierPtr` to protected

`itk::RGBGibbsPriorFilter` still carries a shadow declaration of `m_ClassifierPtr` because `itk::MRFImageFilter` keeps that member `private` and its `SetClassifier` is non-virtual — so the subclass method *hides* rather than overrides, and the two classes maintain independent classifier pointers with divergent `SetClassifier` semantics.

Follow-up to #6254, which removed the two `NumberOfClasses`/`MaximumNumberOfIterations` shadows in the same class but deliberately deferred this one because the fix requires a base-class API change.

Evidence (file:line, verified against main 2026-07-29)

`Modules/Segmentation/MarkovRandomFieldsClassifiers/include/itkMRFImageFilter.h:421` — in the `private:` section that begins at line 384:

```cpp
typename ClassifierType::Pointer m_ClassifierPtr{};
```

`Modules/Segmentation/MarkovRandomFieldsClassifiers/include/itkRGBGibbsPriorFilter.h:216` — the shadow, same type:

```cpp
typename ClassifierType::Pointer m_ClassifierPtr{};
```

Both `SetClassifier` declarations are non-virtual and neither carries `override`:

```cpp
// itkMRFImageFilter.h:248
void SetClassifier(typename ClassifierType::Pointer ptrToClassifier);

// itkRGBGibbsPriorFilter.h:131
void SetClassifier(typename ClassifierType::Pointer ptrToClassifier);
```

Because the methods are non-virtual, dispatch is chosen by the **static** pointer type:

| Static type of pointer | Method invoked | Which `m_ClassifierPtr` is mutated | Throws on `NumberOfClasses <= 0`? | Forwards `SetNumberOfClasses`? |
|---|---|---|---|---|
| `MRFImageFilter *` | base | base's | yes | no |
| `RGBGibbsPriorFilter *` | subclass | subclass's | no | yes |

The bodies are genuinely different, not redundant:

```cpp
// itkMRFImageFilter.hxx
if ((ptrToClassifier.IsNull()) || (m_NumberOfClasses <= 0))
{
throw ExceptionObject(__FILE__, __LINE__, "NumberOfClasses is <= 0", ITK_LOCATION);
}
m_ClassifierPtr = ptrToClassifier;

// itkRGBGibbsPriorFilter.hxx:167-168
m_ClassifierPtr = ptrToClassifier;
m_ClassifierPtr->SetNumberOfClasses(this->GetNumberOfClasses());
```

Subclass `.hxx` references to the shadowed member (all resolve to the subclass copy today):
`itkRGBGibbsPriorFilter.hxx:38` (ctor init), `167`, `168`, `411`, `415` (commented out), `418`, `420`, `542`.

Suggested approach

This is **not** a mechanical shadow removal like #6254; it changes `MRFImageFilter`'s public API and member protection. Order of operations:

1. Make `MRFImageFilter::SetClassifier` `virtual`; add `override` to `RGBGibbsPriorFilter::SetClassifier`.
2. Promote `MRFImageFilter::m_ClassifierPtr` from `private:` to `protected:`.
3. Delete the subclass's `m_ClassifierPtr` declaration. The eight `.hxx` references then resolve to the inherited member.
4. **Decide the semantics deliberately.** Once dispatch is virtual, a call through an `MRFImageFilter *` reaches the subclass override, which skips the base's `NumberOfClasses > 0` invariant and adds a `SetNumberOfClasses` forward. Either the base invariant should be enforced in the override, or the divergence must be documented as intentional. This is the review-blocking design question.
5. Audit for callers that rely on today's static-dispatch behaviour (likely none in-tree, but out-of-tree subclasses of `MRFImageFilter` exist).

Characterization GTests to land **first**, against unmodified code (the test-first protocol used for #6254/#6255):

- `MRFImageFilter.SetClassifierThrowsIfNumberOfClassesUnset` — pins the base invariant.
- `RGBGibbsPriorFilter.SetClassifierViaSubclassUpdatesClassifier`.
- `RGBGibbsPriorFilter.SetClassifierForwardsNumberOfClassesToClassifier` — pins the subclass's extra side effect.
- `RGBGibbsPriorFilter.SetClassifierViaBasePointerCurrentlyHidesSubclassMethod` — documents today's static-dispatch behaviour; update it to assert virtual dispatch after the change.

Related work

- #6254 (merged) — removed the `m_NumberOfClasses` / `m_MaximumNumberOfIterations` shadows in this same class; excluded `m_ClassifierPtr` as invasive.
- #6253 (merged) — `itkVirtual` / `itkFinal` / `itkNonVirtual` macro variants, useful for step 1.
- #6316 (merged) — `-Wshadow-field` enablement, which is what keeps surfacing this class of defect.
- Companion tracker for the rest of the sweep: see the shadow-member sweep issue (this repo).

Start with Modules/Segmentation/MarkovRandomFieldsClassifiers/include/itkMRFImageFilter.h and itkRGBGibbsPriorFilter.h, then inspect the corresponding .hxx references. Add the four named characterization GTests first, including the current static-dispatch behavior. Done means the API change has deliberate invariant semantics, the shadow references are covered, and the relevant tests pass.

Written by the indexing model from the issue text.