Encapsulate increasingly sophisticated boilerplate around running library unit
tests, which we'd been manually synchronizing for three platforms, in a new
run_tests shell function.

Some Boost libraries have multiple libs/$library/test subdirectories with
Jamfiles. On encountering those, we used to emit a warning and skip them
entirely. I was shocked to see that warning for several different libraries in
a local build: I'd had no idea we weren't running those tests. Rework
run_tests logic to loop over every libs/$library/test subdirectory with a
Jamfile.

In fact, rework test logic to generate the list of test subdirectories
*first*, piping them to the run_tests script, which reads from its stdin. This
allows the calling logic to use grep -v to eliminate specific problem test
subdirectories, a finer granularity than skipping all tests for that library.
(It also allows temporary local edits to engage run_tests for some specific
test subdirectory while iterating on a failure.)

Eliminate suppress_tests shell function: we now have a more straightforward
way of bypassing certain specific libraries or tests within libraries.

Introduce sep shell function to produce a separator line within build log
output to help find particular passages.

Replace deprecated Boost.Coroutine with Boost.Coroutine2.