ci: align relocatable package flow with RVS reference

[thananon]

Summary

Brings TransferBench's TheRock-SDK-based package build into closer parity with the ROCmValidationSuite reference flow. Closes the four gaps identified after #262 landed.

Gap 1 — RPATH defaults move into CMakeLists.txt

RPATH list previously lived only in build_packages_local.sh (-DCMAKE_INSTALL_RPATH=…), so plain cmake -DBUILD_RELOCATABLE_PACKAGE=ON .. produced a non-relocatable binary. Defaults now live inside the if(BUILD_RELOCATABLE_PACKAGE) block in CMakeLists.txt. The script drops the three -DCMAKE_*_RPATH flags. The list is also extended to match RVS:

$ORIGIN:$ORIGIN/../lib:/opt/rocm/extras-<MAJOR>/lib:/opt/rocm/lib:/opt/rocm/lib/llvm/lib:/opt/rocm/core-<MAJOR>/lib:/opt/rocm/core-<MAJOR>/lib/llvm/lib

Gap 2 — Auto-computed patch version

VERSION_STRING patch is now derived from git describe --tags --abbrev=0 --match v<MAJOR>.<MINOR>.* followed by git rev-list --count <tag>..HEAD. Falls back to a hardcoded patch (02) when git is unavailable or no matching tag exists. No user-visible change today since no v1.66.* tag exists yet (latest v-tag is v1.62.00); once a v1.66.00 tag is created, builds will be 1.66.<commits-since>.

Gap 3 — RVS-style package release tag

build_packages_local.sh now emits release strings matching the RVS format:

Trigger	Release string
push / schedule / dispatch / local	r<libpatch>.<yyyymmdd> (e.g. r0711.20260504)
pull request	r<libpatch>.<yyyymmdd>.<src-branch>.<commit>
release/* branch	${GITHUB_RUN_NUMBER} (fallback 1)

UTC date captured once at script start so long builds can't straddle midnight. The over-permissive bare rel* branch matcher is dropped (workflow only triggers release/**).

Gap 4 — User-facing TGZ install doc

TGZ install instructions move out of the CI-engineer-facing README_BUILD_PACKAGES.md into a new Sphinx doc at docs/install/INSTALL_TGZ.rst, covering pre-install ROCm, runtime deps (numactl, libnuma1, hsa-rocr — no PCI deps since TransferBench doesn't enumerate via PCI), extraction, PATH/LD_LIBRARY_PATH, and persistent /etc/profile.d setup. install.rst cross-links to it. README_BUILD_PACKAGES.md keeps a minimal CI-maintainer smoke test plus a link to the new doc.

Test plan

• cmake -B build_verify -DBUILD_RELOCATABLE_PACKAGE=ON -DROCM_MAJOR_VERSION=7 … prints -- TransferBench version: 1.66.02 (correct fallback)
• Bash release-tag logic smoke-tested in default / release / PR modes — produces r0711.20260504, 42, r0711.20260504.feature.foo.<commit> respectively
• CI: push triggers build-relocatable-packages.yml; verify package filenames carry the new r<libpatch>.<yyyymmdd> release string
• CI: confirm readelf -d on the built binary shows the extended RPATH list and no $HOME/rocm-sdk/install leak
• Render Sphinx docs (python3 -m sphinx -W -b html …) and confirm INSTALL_TGZ.html builds without warnings

[alex-breslow-amd]

I would try to automate the first-level troubleshooting you listed if possible. Then the doc can just refer the user to the output that's already part of the logs.

[thananon]

@gilbertlee-amd good catch — --help was wishful thinking on my part. Pushed a fix that:

• Replaces TransferBench --help with plain TransferBench in both the README smoke test and INSTALL_TGZ.rst. On develop today, argc <= 1 (Client.cpp:41) prints version + usage + preset list + detected topology then exits 0, which is the right load-time check.
• Adds a forward-looking note in the README that once the help and smoketest presets graduate from candidate to develop, the smoke test should switch to TransferBench smoketest.

Left @alex-breslow-amd's two open suggestions (run a real workload like a2a; auto-run the troubleshooting ldd/readelf) for follow-up — both are good ideas but bigger than this PR's scope (the a2a one in particular wants the smoketest preset to land first).

[alex-breslow-amd]

@gilbertlee-amd good catch — --help was wishful thinking on my part. Pushed a fix that:

• Replaces TransferBench --help with plain TransferBench in both the README smoke test and INSTALL_TGZ.rst. On develop today, argc <= 1 (Client.cpp:41) prints version + usage + preset list + detected topology then exits 0, which is the right load-time check.
• Adds a forward-looking note in the README that once the help and smoketest presets graduate from candidate to develop, the smoke test should switch to TransferBench smoketest.

Left @alex-breslow-amd's two open suggestions (run a real workload like a2a; auto-run the troubleshooting ldd/readelf) for follow-up — both are good ideas but bigger than this PR's scope (the a2a one in particular wants the smoketest preset to land first).

That's fine by me.

[thananon]

Codeql broken unrelated to the PR. If we are ok lets merge this and create a test release branch to verify this workflow end to end.