Update rocm and adds ccache

This updates ROCM to 5.1.3 in the ROCm image and it also adds ccache to speedup the build.

Here is the build time graph for the HIP build:

Note that we achieving a 100% cache hit which is unreasonably high in a normal build, but you get an idea of its build speed up.

Also note that with ROCm 5.1.3 we are not getting this 40GiB memory usage of a single amdclang++ process anymore. With a concurrency of 6 threads we see a much lower usage in total ~10Gib. I have not been watching the whole time the mem usage but at usual bottlenecks it did not increase its memory takes significantly.