[CodingInAVan]

Bumps the version to 1.3.0rc7. This commit touches only CMakeLists.txt GPUFL_VERSION_SUFFIX, pyproject.toml and python/gpufl/__init__.py. release.yml rewrites only the numeric CMake version from the tag, so the suffix has to be committed.

rc7 contains:

• Group a launch's processes and record their identity #176: Group a launch's processes and record their identity.
• Let the session directory's group read the session lock #177: Let the session directory's group read the session lock.
• Report what each deep window actually collected #178: Report what each deep window actually collected.
• Keep the whole target process tree in the trace lifecycle #179: Keep the whole target process tree in the trace lifecycle.
• Drain PM sampling at least three times per timed deep window #180: Drain PM sampling at least three times per timed deep window.
• Add gpufl trace --stacks=on|off and report the stack choice #181: Add gpufl trace --stacks=on|off and report the stack choice.
• Separate AMD dynamic LDS from the static kernel requirement #183: Separate AMD dynamic LDS from the static kernel requirement.
• Estimate AMD occupancy per GPU target and fix CDNA 3 #184: Estimate AMD occupancy per GPU target and fix CDNA 3 (D9).
• Add opt-in stochastic PC sampling on AMD Instinct GPUs #185: Add opt-in stochastic PC sampling on AMD Instinct GPUs.
• Let the agent choose the channels it uploads #186: Let the agent choose the channels it uploads.

#186 and the backend. rc7 changes no backend upload until the NVIDIA toolkit image's client pin moves. On gpufl-backend dev:

• The AMD runtime image runs gpufl trace without --upload. Its Dockerfile says only the separate contained uploader invokes Java, and the application has no upload path.
• The backend's uploaders run the agent JAR themselves and set GPUFL_LOG_TYPES=device,scope,system,sass (DockerArtifactUploader.java:191, GpuArtifactUploader.java:84, AgentUploader.java:62). They don't go through the launcher, so Let the agent choose the channels it uploads #186 doesn't reach them.
• The NVIDIA toolkit image ships the agent for gpufl trace --upload, but it builds gpufl-client at v1.3.0rc6 (.github/workflows/deploy-experiment.yml:88). Once that pin moves, gpufl trace --upload there uploads every channel the agent finds unless it passes --log-types. An exported GPUFL_LOG_TYPES no longer narrows it.

To release: merge with "Create a merge commit", then tag that merge commit v1.3.0rc7. The tag publishes to PyPI and the Jupyter image.

Next: a gpufl-backend PR pins the AMD runtime image (experiment/amd-runtime/Dockerfile, today 80137e35 from #183) to rc7 for the dev qualification in docs/amd-pc-sampling-plan.md.