Add rocm CI - #343

…hirdPartyLibs into feature/paludettomag1/rocm-ci

…heckout

…uild context

…es before docker build

…nk exclusion

…t explicitly in the Dockerfile

- build.sh: single driver for z5tux (cpu/cuda/hip) and LC (matrix, tioga,
  tuo-cpu, tuo-gpu), cmake+ninja by default, --spack for uberenv builds
- CMakeLists: HIP plumbing for raja, chai, caliper and hypre
- Make python explicit in the py-sphinx externals (spack v1.2 deprecation)

Master reworked the docker CI: the per-platform TPL Dockerfiles and
docker/spack.yaml were replaced by parameterized docker/tpl-ubuntu.Dockerfile
and docker/tpl-rockylinux.Dockerfile layered on the geosx/<os> base images from
docker_base_images, driven by SPEC/DOCKER_BASE_IMAGE from the workflow matrix,
plus a new scripts/reproduce_ci.sh for running a CI row locally.

The ROCm work is ported onto that structure rather than merged alongside it:

* docker/tpl-ubuntu-hip.Dockerfile takes DOCKER_BASE_IMAGE (was
  DOCKER_ROOT_IMAGE) and SPEC as build args instead of hardcoding the spec,
  and reads the build context from the readwrite mount like the other TPL
  Dockerfiles, dropping the uberenv named-context workaround. It keeps its own
  Dockerfile because AMD's ROCm image is not a docker_base_images image and
  ships its own toolchain. Also redeclares AMDGPU_TARGET/ROCM_VERSION in the
  final stage so the ROCm ENV defaults interpolate.
* The ROCm row moves into the workflow matrix in master's style. Rows that do
  not build on a docker_base_images image now set DOCKER_BASE_IMAGE_OVERRIDE.
* CA injection handles apt-based images as well as dnf-based ones, which the
  ROCm and Ubuntu CUDA rows on streak2 need.
* scripts/reproduce_ci.sh gains the ROCm row via two optional trailing matrix
  fields (base_image, extra_env); existing rows are unchanged.
* scripts/run-rocm-ci-local.sh is removed, superseded by reproduce_ci.sh.

Conflict resolutions: took master's newer hypre pin and new hypredrive pin;
took master's hypredrive/addr2line/multi-arch-CUDA changes in the geosx package
and kept this branch's cmake_cache_path/cmake_cache_string conversion on top.

Brings spack v1.2.2 (and the Jul 20th 2026 package repo), the unified build.sh
with HIP support in the TPL superbuild, removal of blueos/Lassen/ESSL support,
and Trilinos defaulting off.

Conflicts: kept this branch's spack_setup_clingo=0 alongside the v1.2.2 spack
pins (all three TPL Dockerfiles pip-install clingo themselves); took v1.2.2's
removal of the ppc64le/blueos and ESSL blocks in the geosx package while
keeping this branch's cmake_cache_path/cmake_cache_string conversion.

docker/spack-rocm.yaml was the only spack env still carrying externals copied
from an Ubuntu 22.04 config, which spack v1.2.2 would take at face value since
they are declared buildable:false. Versions verified against
rocm/dev-ubuntu-24.04:6.4.3 and corrected: python 3.10.12 -> 3.12.3,
py-sphinx 4.3.2 -> 7.2.6, openmpi 4.1.2 -> 4.1.6, zlib 1.2.11 -> 1.3,
readline 8.0 -> 8.2, perl 5.34.0 -> 5.38.2, tar 1.34 -> 1.35, xz 5.2.5 -> 5.6.1,
findutils 4.7.0 -> 4.9.0, netlib-blas/lapack 3.10.0 -> 3.12.0, gmake 4.4.1 ->
4.3, cmake 3.28.3 -> 3.28.6, and mpfr 6.0.2 -> 4.2.1 (6.0.2 was never an mpfr
release). It also picks up v1.2.2's external-spec syntax change, moving the
py-sphinx python constraint from an inline '^' into a dependencies list.

docker/rocky-spack.yaml declares xz as a non-buildable external at /usr, but
the image only installed the xz command, not its headers, so /usr/include/lzma.h
was absent. Nothing needed it until the spack v1.2.2 package repo pulled in
file@5.46, whose configure then failed the Rocky builds with:

    configure: error: xzlib support requested but not found

Ubuntu is unaffected: file never enters its dependency graph.

Note the same file still declares xz@5.2.5 while Rocky 8 actually ships 5.2.4;
that entry is shared with the Rocky 9 rows, which do ship 5.2.5, so it is left
alone here rather than split per-distro.

Bash only treats literal VAR=value words as assignments, so the expanded
${extra_env} was parsed as the command name and the ROCm row died with
'ROCM_VERSION=6.4.3: command not found' before docker build ran.