Performance improvement #2: Flags denormalization.

In the performance schema implementation,
in storage/perfschema/,
implemented the following changes.

1) Define pre compiled instance flags.

Before, in pfs.cc, to check if a mutex event is enabled, the code did:
- Get the PFS_mutex_class from the PFS_mutex
- check the PFS_mutex_class::m_enabled flag
- check the flag_global_instrumentation consumer flag.

This costs:
- several memory access to follow pointers,
- several tests

With this fix, a pre compiled flag is maintained:
- PFS_mutex::m_enabled
that indicates if an event on a mutex instance is enabled or not.

A single test is now sufficient in get_thread_mutex_locker().

Because PFS_mutex::m_enabled is a pre compiled value,
it needs to be refreshed each time:
- setup_instruments changes for the mutex class
- setup_consumers changes for the global_instrumentation consumer.

Table setup_instruments and setup_consumers now 
do update all derived flags when a setup value is changed.

Similar changes have been implemented for other instrument type:
- rwlock
- cond
- file
using the same pattern.

2) Define pre compiled table_share flags.

Similar to 1) for PFS_table_share::m_enabled and m_timed.

PFS_table_share::m_enabled is computed from the rules in
performance_schema.setup_objects.

3) Define pre compiled table handle flags.

Similar to 1) for PFS_table::m_io_enabled and related.

4) Packed performance schema internal structures

Reordered members, removed unused members.

5) Removed the setup objects cache.

Using denormalized flags is a "push" design,
where the thread that changes setup_* tables pushes
all changes to internal structures.

The setup object cache was a "pull" design,
where a thread executing instrumented code had to pull
data from different places at runtime,
to maintain internal structures.

The setup object cache has been removed.