Add coverage module

[apcamargo]

This PR adds the coverage module to MAGpurify.

Changes

• The module calls coverm to compute the contig coverages from BAM files.
• I'm using the parameters recommended by the MetaBAT2 developers: --contig-end-exclusion 75 and --min-read-percent-identity 0.97.
• The user can input BAM files created from mapping the reads to the MAG or to the original metagenome. In the later, the module will only select that appear in the MAG.
• [Up for discussion] The user can input multiple BAM files. If so, only the BAM with the largest average coverage will be the used for outlier identification.
• [Up for discussion] Outlier detection is only performed if the average coverage is over 1.0 in at least in one sample.
• [Up for discussion] In the paper it is stated: "contigs removed with read depth +/- 5x the mean read depth of the bin". My interpretation of that sentence is that the average contig coverage is computed and then contigs with coverage greater than 5average or less than (1/5)coverage are flagged as outliers. Is my interpretation correct?
• The optional parameter --max-deviation allows the user to change the stringency of the refinement.

outliers = ((mag_coverage_df / mag_coverage_df.mean()) >= args["max_deviation"]) | (
    (mag_coverage_df / mag_coverage_df.mean()) <= 1 / args["max_deviation"]
)

[apcamargo]

@snayfach I listed three points that I'd like to discuss before merging the PR (indicated with the [Up for discussion] in the post above).

1. I'm not sure if using just a single sample for outlier detection is the best option. Useful information may be ignored. I choose to do this because it seemed to me that you only used a single sample in the paper. Is that correct?
2. I set the sample minimum average coverage to 1.0. If the mean coverage is less than that, outlier detection doesn't happen. What do you think?
3. Is the outlier detection implemented in the same way you did for the paper?

[snayfach]

Re #1 - Yes I only used a single MAG from a single sample for outlier detection. Using multiple samples and looking at covariation of coverage seems considerably more difficult. Would this require having a co-assembly?

Re #2 - A minimum of 1x coverage seems reasonable. This cutoff might be explored a bit more in a paper. Also the deviation from the mean/median might depend on the mean/median. Higher deviation might be expected for bins with lower average coverage.

Re #3 - Yes this is the same way I defined an outlier

[apcamargo]

Re #1 - Yes I only used a single MAG from a single sample for outlier detection. Using multiple samples and looking at covariation of coverage seems considerably more difficult. Would this require having a co-assembly?

Not necessarily. You can map reads from related samples (eg.: replicates) to your MAG.

I finished writing the README and did some testing. Everything seems to be ok. If you fine with the changes, the PR can be merged.