gdb/amdgpu: provide dummy implementation of gdbarch_return_value_as_value

gdb/amdgpu: provide dummy implementation of gdbarch_return_value_as_value

The AMD GPU support has been merged shortly after commit 4e1d2f5814b2
("Add new overload of gdbarch_return_value"), which made it mandatory
for architectures to provide either a return_value or
return_value_as_value implementation.  Because of my failure to test
properly after rebasing and before pushing, we get this with the current
master:

    $ gdb ./gdb -nx --data-directory=data-directory -q -ex "set arch amdgcn:gfx1010" -batch
    /home/simark/src/binutils-gdb/gdb/gdbarch.c:517: internal-error: verify_gdbarch: the following are invalid ...
            return_value_as_value

I started trying to change GDB to not force architectures to provide a
return_value or return_value_as_value implementation, but Andrew pointed
out that any serious port will have an implementation one day or
another, and it's easy to add a dummy implementation in the mean time.
So it's better to not complicate the core of GDB to know how to deal
with this.

There is an implementation of return_value in the downstream ROCgdb port
(which we'll need to convert to the new return_value_as_value), which
we'll contribute soon-ish.  In the mean time, add a dummy implementation
of return_value_as_value to avoid the failed assertion.

Change-Id: I26edf441b511170aa64068fd248ab6201158bb63
Reviewed-By: Lancelot SIX <lancelot.six@amd.com>