This patches the clang-offload-bundler tool to add a compression option
from a more recent version of clang.
This compression option reduces the size of ROCm's fat binaries.
Those binaries contain .hip_fatbin sections with GPU-specific code,
for each target.
Compression is automatically turned on for all produced outputs
via a wrapper, because it's difficult to identify all the places
where the -compression argument would be needed.
Once upsteam introduces handeling for this argument,
we should drop the wrapper again.
This transistion will create inconsistsency, but I do not think that
it will impact any actual users and it's what's practical to implement.