Split the bulk of the code in LcovReporter.lcov_file out into two free helper
functions, lcov_lines and lcov_arcs.  This is easier to read and will make it
easier to do future planned changes in a type-safe manner.

No functional changes in this commit.

The branch field of a BRDA: record can be an arbitrary textual label.
Therefore, instead of emitting meaningless numbers, emit the string
“to line <N>” for ordinary branches (where <N> is the arc destination
line, and “to exit” for branches that exit the function.  When there is
more than one exit arc from a single line, provide the negated arc
destination as a disambiguator.

Thanks to Henry Cox (@henry2cox), one of the LCOV maintainers, for
clarifying the semantics of BRDA: records for us.

Quite straightforward: a function has been executed if any of its region’s
lines have been executed.

Should fix the test failures with pypy pretending to be python 3.8.

There is a bug somewhere, in which if we collect data in --branch mode under
PyPy 3.8, regions for top-level functions come out of the analysis engine with
empty lines arrays.  The previous commit prevented this from crashing the lcov
reporter; this commit adjusts the tests of the lcov reporter so that we expect
the function records affected by the bug to be missing.

I don’t think it’s worth trying to pin down the cause of the bug, since Python
3.8 is approaching end-of-life for both CPython and PyPy.

.

Doesn’t quite work yet, see discussion in nedbat#1850.

This replaces the custom fn_coverage_missing_in_pypy_38 logic that was
used in earlier commits.

The lcov output tests that are affected by bugs in PyPy 3.8, fail with the
current version of PyPy 3.8 (7.3.11), unlike the other tests annotated with
@xfail_pypy38.  Split this decorator into two variants, xfail_older_pypy38
(used for all the tests that were labeled xfail_py38 prior to this patchset)
and xfail_all_pypy38 (used for the lcov output tests).