The optimisation in write_remoting_arg for sizing, marshalling,
unmarshalling and freeing ref pointers to base types directly was
broken in d458a599 and caused
unnecessary calls to pointer marshalling functions.

The same was true for a similar optimisation for simple structures
where their size can be pre-calculated, freeing omitted and the
NdrSimpleStructMarshall/Unmarshall functions called directly instead
of calling NdrPointerMarshall/Unmarshall first.

Fix this by looking at the type of the referrent instead of the type
of the pointer, making sure to not classify user types as simple
structures.

Additionally remove some dead code that would never be executed
because the code is in an RPC_FC_RP case, inside a !is_ptr if block
(and RPC_FC_RP is a pointer type).