gclosure: Optimise g_closure_[ref|unref]()

Using the CHANGE_FIELD() macro to implement reference counting leads
to poor code generation. This is due to -

* the fields of GClosure are defined as bitfields
* using volatile bitfields forces extra loads

Create specialised macros that operate on short (15bits for ref_count
and 1 bit for meta_marshal_nouse) and use them to implement
g_closure_[ref|unref](). The macros make use of the recently
introduced g_atomic_int16_[add|dec_and_test]() atomic helpers.

The existing asserts in place ensure that the value of ref_count
doesn't spill over into the additional bit used for atomic
instructions (meta_marshal_nouse).

8 jobs for closure/ref_count/devel in 15 minutes and 35 seconds (queued for 7 seconds)