Implement multidimensional scaling approach in regular MMM (#1862)

* feat: implement multidimensional scaling approach in regular MMM

This commit transforms the regular MMM to use the same scaling approach as
the multidimensional MMM, where data is stored in original scale and scaling
is applied within the model graph using PyTensor operations.

Key Changes:
- Add scaling configuration parameter to MMM constructor
- Store original scale data in preprocessed_data instead of pre-scaling
- Compute scale factors (target_scale, channel_scale) and store as pm.Data variables
- Apply scaling transformations within the model graph
- Fix sample_posterior_predictive to return original scale data by default
- Update transformation logic throughout codebase to work with original scale data
- Add original scale contribution variables (channel_contribution_original_scale, etc.)
- Update channel_contribution_forward_pass to use computed scale factors
- Preserve backward compatibility with existing transformer approach

Fixes:
- Resolves KeyError when accessing y_out_of_sample['y'] from sample_posterior_predictive
- Ensures out-of-sample predictions return same scale as training data (original scale)
- Fixes plotting methods to work correctly with original scale variables

Tests:
- 442/444 tests pass (2 implementation-specific tests marked as xfail)
- Add fallback logic for backward compatibility
- Update test expectations for new scaling behavior

This change ensures consistency between regular MMM and multidimensional MMM
scaling approaches while maintaining API compatibility.

* pre-commit

* changing a few things

* Adding tests

* Modifying to nc

* Solve test issue

Author: cetagostini

docs/source/notebooks/mmm/mmm_allocation_assessment.ipynb

@@ -4152,7 +4152,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.10.18"
+   "version": "3.12.11"
   }
  },
  "nbformat": 4,