We've seen at least one incident in the field that suggests we
used the same identifying token in back to back operations, which
breaks certain invariants in the asynchronous-completion bookkeeping.
Harden against this by making sure that we don't reuse tokens in
nearby proximity. This is done by making the low 8 bits a sequence
count, i.e. guaranteed unique within the last 256 operations, while
keeping the upper bits random so that apps can't guess at correct
tokens.
Fix 63691912
Fix 63553575
Test: CTS backup pass to completion plus manual inspection
Change-Id: I321c3a2b3f4203836800bb72db7332bd82e54aaf
6935755f43