Removing generation of matrices which can't be modeled

[amcim]

Two changes were made to generate_matrices_ml.R code. The motivation is the number of matrices that are created but did not have enough of the minority class, enough genomes, etc. This PR stops the attempt to model these.

• Skip unusable derived matrices: New .skipUnusableMatrix(), called from .parquet2LOOMatrix, .parquet2LOODrugMatrix and .parquet2CrossDrugTestMatrix. These matrices are built by removing genomes, so they can end up tiny or single-class. There is no point modeling those, so they're now skipped with a logged reason instead of written.
• Run the balance check after the join, not before: matrix is built by joining metadata (drug phenotypes) to the feature tables (gene counts) on genome_id. Some genomes are in metadata but have no feature rows, so the join drops them, but skipImbalancedMatrix was running beforehand and counting genomes that would never make it in. The check now sees the genomes that survive the join.

[AbhirupaGhosh]

The scripts ran without any error. I made some minor changes I observed in downstream functions.