feat: [AMD Support via SemiAnalysis][4/4] Optional job observation and diagnostics

[cquil11]

"Nice to haves" such as job observation and diagnostics, stacked above the complete AMD/MoRI/ATOM implementation. The former required-runtime fixes have moved into #431.

• srtctl wait streams the allocation log and returns its actual terminal exit status. An empty queue never counts as success; transient observation failures do not resubmit or cancel jobs.
• Decode non-UTF-8 failure logs without losing useful diagnostics.
• Capture SGLang Router and MoRI package versions in runtime provenance.

These changes do not enable GPUs, allocate workers, route traffic, or configure KV transfer. Native apply and job execution do not require this PR. The InferenceX CI launcher currently uses wait for observation, so its validation pin includes this optional layer.

Review the stack

The upstream PRs target NVIDIA main and are manually stacked, so later upstream diffs include preceding layers. The downstream fork PRs form an actual branch-based stack (main → #1 → #11 → #7 → #14); each fork PR's Files changed tab isolates its layer for easier review.

• AMD hardware and cluster runtime: upstream #431 ↔ fork #1.
• vLLM/MoRI: upstream #432 ↔ fork #11.
• ATOM/AToMesh: upstream #433 ↔ fork #7.
• Optional observation and diagnostics: upstream #476 ↔ fork #14.

Merge in that order, then rebase the remaining layers. ATOM uses Mooncake, not MoRI; its stack position is review ordering, not an engine dependency. The core runtime does not require the observation layer, but the InferenceX launcher uses its srtctl wait command.

E2E validation is maintained only in the base PR: upstream #431 / fork #1.