Before this fix, the following query:
  SELECT * FROM performance_schema.status_by_account
could crash the server under load.

The root cause is that the monitoring query inspects
all running sessions (THD) in the server,
and reads data (without any locks) from members of THD while the session
itself is running, creating race conditions.

In particular, while the THD object itself is guaranteed
to exist in memory, members of THD that contains pointers
to other memory fragments, such as
  THD::m_main_security_context::m_user::m_ptr
may change while the session represented by THD is running.
Reading these members directly is not safe,
and caused crashes in pfs_visitor.cc, in:
- match_host()
- match_user()
- match_account()

The fix is to abandon inspecting THD::m_main_security_context,
and inspect instead attributes of PFS_thread to find out if a
given THD belongs to a given account / user / host.

To achieve this, it is necessary to find the PSI_thread associated
with a THD.

This relation existed in the code already,
with member THD::scheduler:m_psi.

The performance schema instrumentation stored in the thread scheduler
has been moved into THD directly (THD::m_psi), to make the relationship
usable independently of the thread scheduler.

Note that THD::m_psi is now maintained using atomic operations,
which allows to safely read THD::m_psi of a running THD session
from a monitoring query executing
  SELECT * FROM performance_schema.status_by_account