* Only IO endpoints calling back into scheduler so far
* IO::Buffer totally stubbed out
* Organization of utility methods likely to change

For test_io_buffer.rb:test_inplace_operators

This also necessitated implementing initialize_copy for the dup'ed
buffer in the test, along with to_s, inspect, and hexdump for when
there's an error condition.

Moving to Java 13+ would allow us to do absolute puts without the
position changing.

Position is not part of IO::Buffer so we must constantly reset it.

Pushing toward the blessed path always receiving a ByteBuffer, so
we can avoid recreating it constantly for small reads.

This unfortunately does not work for most files in JRuby, since we
usually use a real native file descriptor on POSIX platforms. This
will need to be expanded to work with a native fileno but produce
a MappedByteBuffer.

This doesn't pass test_string_mapped because we do not lock
String objects like MRI does and cannot raise an error when the
String owned by the Buffer is modified independently.

This cleans up how blocking works and fixes small parts of the
scheduler dispatch logic. Fibers start out nonblocking by default,
which means entering such a fiber puts the system in a schedulable
state (decrements root fiber's default blocking count of 1 to 0).
While schedulable, blocking calls will redispatch into the
scheduler to give it a chance to handle the blocking operation and
optionally transfer to another unblocked fiber.

The handling of blocking and unblocking will require some
additional work, since currently only a fiber yielding can unblock
its parent, but we are getting the pieces wired up.

All of CRuby's test/fiber/test_scheduler.rb passes with these
changes.

It still all passes 🎉

Working on failures in mri/fiber/test_io_buffer.rb

* This includes two PRs to improve behavior and testing:
  * ruby/ruby#8728
  * ruby/ruby#8729