feat(runners): add GMI Cloud launchers (H100/H200/B200/B300/GB200)

[OCWC22]

Upstream PR draft — do not yet send

Title: runners: add GMI Cloud as a first-class launcher (H100/H200/B200/B300/GB200) — closes B300+ NeoCloud gap

Source: /Users/chen/Projects/Inferscope/prompt-exports/2026-04-27-inferencex-gap-analysis.md (§6.3)

This PR adds GMI Cloud as a supported runner across the full Hopper +
Blackwell range, mirroring the existing cw/cr/dgxc/nb/nv pattern.

What changes

• New launchers: launch_h100-gmi.sh, launch_h200-gmi.sh, launch_b200-gmi.sh,
  launch_b300-gmi.sh, launch_gb200-gmi.sh
• Promoted existing fork-only docs to upstream:
  runners/GMI_QUICKSTART.md, GMI_QUICKSTART_H100.md, GMI_QUICKSTART_GB200.md
• No methodology changes; identical CSV output to existing launchers
• No changes to benchmarks/, utils/, or .github/workflows/

Why

• GMI Cloud is a NVIDIA Dynamo 1.0 launch partner with bare-metal access
  to the full Blackwell + Hopper range
• Currently the only B300/GB200/GB300 launchers are launch_*-nv.sh
  (NVIDIA-direct); this PR closes the "B300+ outside NVIDIA's own infra"
  gap. CoreWeave/Crusoe/Nebius launchers max at B200.

Validation

• Tested against dsv4_fp4_b200_vllm.sh on GMI B200 (cell-name: )
• Tested against dsv4_fp8_h200.sh on GMI H200 (cell-name: )
• Multi-node disagg recipe disagg-gb200-1p1d-dep8-tep8.yaml validated
  on GMI GB200 NVL72

Diagnostic add-on (separate companion blog, not this PR)

Touchdown Labs is publishing an inferscope.ai post citing InferenceX
as the substrate, using our open-source inferguard-cli to surface
PD-disagg failure modes during the multi-node DSv4 GB200 sweep. Linked
in the post; no code dependency on this PR.

Validation status (filled in after smoke runs)

• H200 smoke validated: cell SHA
• B200 smoke validated: cell SHA
• B300 smoke validated: cell SHA
• GB200 multi-node smoke validated: cell SHA

Send when

All 4 boxes above are checked, plus the TDL companion blog draft is locked.