rocm rebuild is like 30 mins

not sure --link will work here

could probably avoid this if llama.cpp were subtree instead of submodule, avoiding need to copy into the image build stuff that changes all the time like .git

it still compiles and runs ok on 7900xt

performance tests showing no significant difference, indicating the thing works as good as it did before the changes

initial smokechecks showing successful rocm amd64 compile + equivalent performance in smoke tests