feat: support multinode for JSON output (rebase of #54 onto v0.10)

[jingxu9x]

Rebases #54 onto current main (v0.10) and resolves the conflicts that made it un-mergeable. Original work by @shadychan (credited via Co-authored-by); this supersedes #54.

Why

On multinode --json runs, output->recordDevices(worldSize) reached JsonOutput::recordDevices, which iterated the argument as a local device count and called cuDeviceGet past the local GPU count -> CUDA_ERROR_INVALID_DEVICE. #54 fixed the JSON multinode path but went stale against v0.10, which meanwhile unified the multinode/single-node output paths and added --pair/UUID handling plus a per-node "more processes than GPUs" warning.

What changed vs #54

• Rebased onto v0.10; kept all of main's newer logic (--pair/UUID selection, the per-node oversubscription warning). Refactored printGPUsMultinode / printGPUs into exchangeDeviceInfo + a single recordDevices.
• Uses main's gethostname() / env->getHostname() (the localHostname symbol feat: support multinode for json format output #54 relied on no longer exists on main).
• Fixes the root cause: the sole recordDevices caller now passes deviceCount instead of worldSize, so device listings iterate the actual device count. PerfParserOutput reports worldSize under multinode (participant count), deviceCount otherwise.
• Preserves main's runtime worldSize > 1 dispatch inside the #ifdef MULTINODE branch, so a MULTINODE build run with one rank still uses the single-node listing (avoids getPaddedProcessId(0) / log10(0)).
• Header cleanup: recordDevices(int deviceCount); exchangeDeviceInfo declaration guarded by #ifdef MULTINODE.

Testing

Validated on real hardware (GB200/GB300 multinode) — the multinode --json path that previously failed with CUDA_ERROR_INVALID_DEVICE now completes and emits per-rank device info. Single-node and single-rank MULTINODE runs behave as before. Please re-run your CI / multinode validation to confirm on your side.

Known limitations (pre-existing on main, not addressed here)

These exist on main today and are out of scope for this rebase; flagging for visibility:

• PerfParserOutput does not assign per-rank localDevice in multinode (device assignment is coupled to the text/JSON output helpers; the natural owner is MultiNodeEnv).
• --pair UUID validation is per-rank and non-collective: a rank returning early on a missing UUID can leave peers blocked in MPI_Allgather.
• The JSON warning field is now an array (from feat: support multinode for json format output #54) to retain multiple per-node warnings, a schema change from the previous string. Happy to gate this behind a compatibility decision if you prefer.

Co-authored-by: Shady Chan chenyulin.shady@bytedance.com

[jingxu9x]

@esitaridi can you help review this pr?