fix adaptors

[clemsgrs]

Summary

Fixes four issues in vision adaptors:

• KNNRegressor.fit incorrectly unpacked the return value of preprocess_features, which returns a single array.
• Non-survival LinearProbingRegressor and MultiLayerPerceptronRegressor passed [B] labels to MSELoss against [B, 1] model outputs, causing unintended broadcasting.
• Survival LinearProbingRegressor and MultiLayerPerceptronRegressor computed cumulative survival along dim=0. During training, logits are [num_shots, nbins], so this ran the cumulative product
  across shots instead of survival bins. This now uses dim=-1.
• 2D segmentation inference stitched patch predictions with swapped axes. Task 9 image-size is [width, height], while NumPy masks are [height, width]; inference now allocates (height, width),
  writes patches with [y, x] indexing, and no longer transposes the final mask.

Impact

• the KNNRegressor bug can cause fitting to fail for 3+ shots, or behave incorrectly with exactly 2 shots.

• the non-survival MSE broadcasting issue is serious for non-survival regression because it changes the objective from per-sample regression to effectively pushing all predictions toward the batch mean label. However, the practical impact should be limited as our sole regression task is a survival prediction task, so we have not used these non-survival regression heads.

• the survival axis issue affects the training objective for survival regression heads. Prediction did not generally crash because case-level test embeddings are passed one case at a time as a 1D vector, where dim=0 happened to be the bin axis. Training, however, uses all shots as [num_shots, nbins], so the survival/censoring part of the loss was batch-order dependent.

• the 2D segmentation issue affects segmentation-upsampling for Task 9. On square-ish data this can be partially masked by the final transpose, but local patch content is transposed and non-square images or edge patches can be cropped/placed incorrectly.

[clemsgrs]

@copilot review the changes in this PR

[Copilot]

@copilot review the changes in this PR

I've reviewed the changes in the last 4 commits. No significant issues, bugs, logic errors, or security vulnerabilities were found in the updated code.