Background:
GTID sets are stored internally using class Gtid_set, where
intervals of Gtids are stored using linked lists.
The member function Gtid_set::to_string returns a string representation
of the set. The member function Gtid_set::get_string_length returns the
length of the string, without actually computing the string (this is useful
so that a caller can allocate a buffer before computing the actual string).
Gtid_set::get_string_length caches the string length, so that if
get_string_length is called twice, the length does not need to be recomputed.
The cached string length is forgotten when the set is modified.

Problem:
There was one case where we had forgotten to forget the cached string length:
Gtid_set::clear(). This could cause get_string_length() to return a too big
number after calling Gtid_set::clear(). This in turn would mean that
Gtid_set::to_string would allocate more than one byte, set the first byte
to '\0', leave the rest uninitialized, and return a string length that is
longer than 0.

Since the generated string is NULL-terminated this did not cause much trouble,
except in one place: in 5.7,
performance_schema.replicate_connection_status/RECEIVED_GTID_SET stores a
Gtid_set in a BLOB column. Since it is a BLOB, it does not count '\0' as
end-of-string. So it would print all the allocated characters, including
uninitialized bytes.

Fix:
Clear Gtid_set::cached_string_length in Gtid_set::clear.