Fix RR-ClArC to freeze layers before CAV layer during fine-tuning

[Copilot]

Problem

RR-ClArC's fine-tuning procedure was incorrectly optimizing all model parameters, when it should only update layers from the CAV layer onwards. According to the original paper and as spotted by @kasia284, the method should freeze all layers before the layer where gradient reprojection is applied during fine-tuning.

Solution

This PR modifies the apply_model_correction() method in RRCLARC to:

1. Identify layer ordering - Determine which layers come before the specified cav_layer in the model's execution order
2. Freeze early layers - Set requires_grad=False for all parameters belonging to modules before cav_layer
3. Selective optimization - Update the optimizer to only include trainable parameters (from cav_layer onwards)
4. State restoration - Restore the original requires_grad state after fine-tuning completes

Example

For a model with layers [layer0, layer1, layer2, layer3] and cav_layer='layer1':

• Before: All layers were updated during fine-tuning
• After: Only layer1, layer2, and layer3 are updated; layer0 is frozen

Implementation Details

The freezing logic properly handles:

• Nested modules - Parameters like layer0.conv.weight are correctly frozen when layer0 should be frozen
• Edge cases - Top-level parameters are handled appropriately
• State management - Original requires_grad values are restored after training to avoid side effects

Testing

Comprehensive tests verify:

• ✅ Frozen layers are not updated during fine-tuning
• ✅ Trainable layers are correctly updated
• ✅ Works with different CAV layer positions (first, middle, last layer)
• ✅ Handles nested module structures correctly
• ✅ Original parameter states are properly restored

Impact

• This fix is specific to RR-ClArC only and does not affect other ClArC variants (A-ClArC, P-ClArC)
• The change aligns the implementation with the original paper specification
• Expected to have minor practical impact since the default configuration uses only 1 fine-tuning epoch

Fixes the issue reported by @kasia284 regarding incorrect layer freezing in RR-ClArC fine-tuning procedure.

Original prompt

---

This section details on the original issue you should resolve

<issue_title>RR-ClArC should freeze layers up to l-1 while finetuning</issue_title>
<issue_description>Full credit for spotting this to @kasia284.

In RR-ClArC the fine-tuning procedure should freeze all layers before the layer on which we apply the gradient reprojection. It's probably a minor error, because there's only one fine-tuning epoch, but still definitely worth fixing.</issue_description>

<agent_instructions>Check out the original paper, double check that the code you change doesn't interfere with other important execution traces!</agent_instructions>

Comments on the Issue (you are @copilot in this section)

Fixes #23

---

✨ Let Copilot coding agent set things up for you — coding agent works faster and does higher quality work when set up for your repo.