Make mse work for dense predictions

[rcannood]

Describe your changes

file_prediction.yaml asks for a double normalized layer -- it says nothing about sparsity. Every method we ship happens to write a csc_matrix, so this has never bitten, but a method returning a dense array crashes the metric:

>>> (sol.layers["normalized"] - pred.layers["normalized"].toarray()).power(2)
AttributeError: 'matrix' object has no attribute 'power'

sparse - ndarray yields a np.matrix, which has no .power(). A contributor hitting this gets a crashed metric instead of a poor score, which is the wrong failure mode -- especially with errorStrategy = 'ignore' turning it into a silently missing row.

Coercing both layers to CSR first. Scores are unchanged for the sparse case:

sparse prediction: (1.2189312040309257, 0.9132953464791143)
dense  prediction: (1.2189312040309257, 0.9132953464791143)

Part of a series of PRs coming out of a pre-run review of the benchmark.

Checklist before requesting a review

• I have performed a self-review of my code

• Check the correct box. Does this PR contain:

  • Breaking changes
  • New functionality
  • Major changes
  • Minor changes
  • Bug fixes

• Proposed changes are described in the CHANGELOG.md

• CI Tests succeed and look good!