Fix NdrFreeBuffer to use the fBufferValid flag to determine whether or 
not I_RpcFreeBuffer needs to be called.

Reverse the order of the release statements in NDRSContextMarshall2 for 
the case where the context handle doesn't have valid data so that the 
release that releases the lock comes first and then the optional second 
release doesn't need to release the lock.

Release them after processing of a request has finished to avoid a slow 
memory leak if the association isn't released for ages.

wined3d: Only store the palette index in the alpha component when the primary render target uses P8.

This was already done in other places but I missed this spot.