Exclude additional regions from training data

[petermchale]

In the germline case, in order to test the following (non-coding) positive controls, we must remove them from the list of putatively neutral regions upon which we train the model.

• Confirm that predicted constraint stratified by collections of genomic elements (e.g., promoters, enhancers, TAD boundaries) mimic corresponding results from other studies (e.g., Figure 2 from https://www.nature.com/articles/s41588-018-0062-7)
• Confirm that high constraint is seen in ultraconserved non-coding elements a là page 4 of https://www.nature.com/articles/s41576-021-00424-x
• Confirm that known essential, noncoding genes such as tRNAs are constrained