feat(runtime): expand SGLangConfig to kitchen-sink (B3 followup to #974)

[itsmiso-ai]

Followup to #974 (B2 — initial SGLang runtime backend). Once we have
real workload data on which SGLang flags operators reach for, expand
SGLangConfig to cover the remaining SGLang CLI surface. Candidates:

• LoRA hot-loading (--lora-paths runtime updates) — needs a separate
  CRD API for dynamic adapter registration
• Speculative decoding perf tuning
  (--speculative-accept-threshold-single,
  --speculative-accept-threshold-acc)
• Data-parallel rendezvous / multi-replica orchestration
• PodMonitor for SGLang's /metrics endpoint (scraped today but not
  wired into HPA defaults)
• Remaining minor flags: model override (--model), reasoning-content
  (--reasoning-content), return-logprob (--return-logprob), log
  level (--log-level), trust-remote-code, skip-tokenizer-init, etc.
• A typed LoRAConfig (vs the current []string for loraModules)
  with named adapters

Filed by the same operator running LLMKube's Foreman on a self-hosted
fleet. Will coordinate with the maintainers on which subset is highest
value for the foreman workload.

[itsmiso-ai]

Closing this out in #1103. Five of the six candidates (A, B, C, D, E) shipped in that PR; the sixth (F, multi-replica DP rendezvous) is deferred to the architectural follow-up in #1102 because a clean fix needs a StatefulSet + headless service path or a companion CRD — neither is reviewable as a side effect of this PR.

Summary of what landed in #1103:

• A — --model, --reasoning-content, --return-logprob, --log-level, --trust-remote-code, --skip-tokenizer-init as typed SGLangConfig fields.
• B — SGLangSpeculativeConfig.AcceptThresholdSingle / AcceptThresholdAcc + validator.
• C — typed SGLangLoRAAdapter{Name, Path} plus the legacy LoraModules (deprecated in godoc). Unified helper normalizes both forms to --lora-modules name=path[,name=path]; typed wins on collision.
• D — the PodMonitor + HPA wiring (feat(runtime): add SGLang runtime backend (#974) #1059) was already in place; PR feat(runtime): close out the SGLang kitchen-sink (#1060): minor flags, accept thresholds, typed LoRA adapters, LoRAAdapter CRD #1103 just documents it under the SGLang section of model-matrix.md and the metrics-driven-autoscaling guide so the sglang:num_running_reqs default is discoverable.
• E — new namespaced LoRAAdapter CRD + reconciler that drives SGLang's /v1/lora_adapters/{load,unload} HTTP API. Finalizer-driven unload on delete. Mount path is operator-provided in InferenceService.spec.extraVolumes; the controller does not auto-mount.

make test, make lint, and make lint-all (cross-GOOS) all green on my side; full controller suite + make check-helm-rbac clean. Review-bounded to runtime_sglang*.go and the new loraadapter_controller.go.