Store GLM-PCA factors as LinearEmbeddingMatrix in SingleCellExperiment

[willtownes]

As a user I can apply dimension reduction to a SingleCellExperiment object and access both the factors and the loadings from a LinearEmbeddingMatrix in the reducedDims slot.

• Verify the factors matrix U is stored properly
• Verify the loadings matrix V is stored properly
• Verify there is not an implication in the SingleCellExperiment documentation that UV is supposed to be the mean of the data.
• Make clear in the documentation that UV is not a predicted mean of the data.

[willtownes]

note: the offsets, link function, and any row- or column-wise covariates are all needed to be able to make predictions for the mean of the data. Should these be stored in the metadata of the SingleCellExperiment to facilitate prediction? Should we write a predictGLMPCA method? This would be useful for things like molecular cross-validation where they need to reconstruct a "denoised" version of the original data matrix.

[kstreet13]

I think that if we switch to storing GLM-PCA results as a LinearEmbeddingMatrix, then we should try to keep all the output in that object. So the offsets, link function, etc. could be stored in the metadata slot of the LinearEmbeddingMatrix, rather than the metadata slot of the SingleCellExperiment.