This code was trying to block negative values of memid (incorrectly).
I have run more tests and found that it's possible to set some negative
values (e.g. on TKIND_INTERFACE: -1, -4..-8, -500..?). Unless we know
what values are allowed it's probably better to not check memid.

wined3d: Implement SetDestWindowOverride() by simply making the context current on a different window.

The swapchain's current window might be changed by SetDestWindowOverride().

Clean valgrind runs of tested types and comparison with MIDL-generated
code suggests that the existing code written in the finally block for
client code is sufficient.

Remove the difficult manual walking of the alias tree and make the
code easier to read and less likely to break if a new type is added.

Use a switch statement instead of multiple if statements and
explicitly write out the fixed array cases.

widl: Remove some FIXME comments and replace them with comments explaining why the current behaviour is correct.

widl: Fix the writing of expressions in un-typedef'd structures by using write_type_left to write out the type, which adds the "struct ..." if necessary.

Also improve the type comparison code to not crash on unnamed
structures (although further changes need to be made to fix the output
in this case) and make it more efficient by first doing a comparison
on the values of the type pointers.

The FC code requires more than 4 bits to represent it and the
correlation descriptor and union formats only allow 4 bits to
represent the type referred to.

This allows write_conf_or_var_desc to have the context enabling it to
write out an appropriate correlation descriptor when iid_is is used
from within a structure.

widl: Use typegen_detect_type to determine which types should be written for the pointer description in write_struct_tfs.

Otherwise extra pointers for user types could be written.

widl: Don't use the return value of write_embedded_types to determine whether a type contains pointers or not.

It isn't correct in all situations, so type_has_pointer has to be used
anyway so using the return value of write_embedded_types just clutters
the code.

Avoid duplicating the pointer writing code.

The check for whether a pointer is the last one is redundant since
typegen_detect_type does a stricter check on the ref type. Replace one
other instance of last_ptr with a more readable equivalent.

Change write_pointer_tfs to rely on the detected type of the type to
determine when to write a simple pointer rather than relying on
typestring_offset not being set anywhere else.