For record_scheduler_t, context switch on encoding boundaries instead
of instruction boundaries to avoid splitting up encodings from their
instructions.

We end up not supporting unread_last_record() for record_scheduler_t
as we do not have the pre-prev record to identify a first encoding in
a sequence.

Adds a unit test.

Inserting synthetic tid,pid records so the switches are actually
visible will be added separately.

Issue: #6635

…ull into this separate PR)

…cord-sched-encode-bound