tiny-plot: legend sorted by natural order, other reliability improvements

[AlexTate]

Scatter plot legends are now sorted by natural order while preserving the intended point layer order (groups with fewest points are plotted on top of the stack).

This PR also includes a number of reliability improvements:

• Fixed an edge case that occurs when the outgroup has all zero counts in one or both conditions. Previously this would result in the outgroup having an entry in the legend without any points being plotted
• Fixed an edge case that occurs when there is no outgroup, and the first group has all zero counts in one or both conditions. Previously this first plotted group wouldn't be recognized as a zero count group, so it would have an entry in the legend without any points being plotted
• Properly handling an empty scatter plot when neither the outgroup nor the other groups can be plotted due to one or both conditions having all zero counts
• Fixed calculation of z-order to avoid group-outgroup conflicts and group-line conflicts. Fortunately the way matplotlib breaks z-order ties always worked out in our favor so this issue was never expressed. Now it's proper.
• Added helpful error messages so that scatter_dge plots aren't silently skipped if inputs are empty (due to class filters, etc.)

Additionally, the sorted_natural() function in tiny.rna.util now accepts a key= parameter just like Python's sorted() does

Closes #285

[taimontgomery]

Tested successfully with ram1 data.

[AlexTate]

Exception handling in tiny-plot was also improved so that it is more useful during non-standalone runs.

For each worker exception:

• The tracebacks are printed to stdout so that they end up in the logfile produced by cwltool

At the conclusion of plotting:

• The replot instructions are printed only once regardless of the number of workers that failed
• A summary of the exceptions organized by plot type is printed (no noisy tracebacks)
• The user-friendly message + summary is printed to stderr so that users see it during pipeline runs

Exception handling is also consistent regardless of sequential/multiprocessing or debug mode