PERFORMANCE_SCHEMA.THREADS.PROCESSLIST_STATE

Prior to this fix, the server code executed the following calls to the 
performance schema for each change in processlist state:
- pfs_set_thread_state_v1()
- pfs_start_stage_v1()

pfs_set_thread_state_v1 sets the THREADS.PROCESSLIST_STATE column with a 
string. To protect the string pointer, a PFS_lock is used, which causes two
atomic operations per call.

pfs_start_stage_v1 sets the current thread stage key.

The problem is that using atomics causes overhead, and that setting both a 
state string and a stage key is redundant, since the thread state string is
in fact coming from the stage.

This fix simplify the instrumentation:
- the call to pfs_set_thread_state_v1() is removed
- the implementation of column THREADS.PROCESSLIST_STATE now extract the 
  state string from the stage provided by pfs_start_stage_v1.