Before this fix, when sending a massive load spike to a freshly started
server, the server would allocate a huge amount of memory,
possibly leading to Out Of Memory failures.

The root cause is how concurrent request to grow a scalable buffer
are handled, for the performance schema.

Before the fix, N threads could run in parallel each allocating
memory, with later N-1 threads freeing the memory allocated
when a collision is detected after the fact.

With the fix, N threads attempting to expand a buffer in parralel
will block and wait, with only one thread performing the allocation,
so that collisions are prevented.

See comments added in PFS_buffer_scalable_container::allocate()
for details.