[12.9.x] Fix host-function callbacks in replayed graphs

[Andy-Jost]

Description

Manual backport of #3061 to 12.9.x, addressing #790. CUDA graphs containing a host function could crash on their second launch because the bindings routed callbacks through trampoline state that was freed after the first invocation. driver.cuLaunchHostFunc and runtime.cudaLaunchHostFunc now pass the callback and user data to CUDA unchanged.

• Synchronize the host-function blocks regenerated for CUDA 12.9 after confirming their unfixed generated baselines exactly match the maintenance branch. Preserve the existing runtime Tempita layout.
• Validate callback values, stream, status, and invocation count after synchronization, and exercise five launches of each captured graph. The driver test creates and destroys its own context for this branch.
• Document callback and user-data lifetime requirements and add the release note.

The CUDA 12.9 source build and simulated callback/replay checks passed. The GPU regression cases could not be rerun because the GPU test host was unavailable.

Checklist

• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[mdboom]

As a backport of the generated stuff, this looks correct. I'll defer to @juenglin's review of #3061, and if that gets a ✅ , then consider this approved as well.