[Feature]: Persist per-model LM Studio load configuration when PAIR manages llmster

Feature request

Please support persistent, per-node/per-model LM Studio load options when PAIR manages llmster.

Current behavior

The current LM Studio engine manifest loads a model with only:

lms load <model>

PAIR's UI supports downloading, loading, ejecting, and deleting models, but does not expose or persist important LM Studio load parameters.

Use case

Our existing GPU workstation requires deterministic settings such as:

lms load qwen/qwen3.6-35b-a3b   --gpu max   --context-length 200000   --parallel 2   --identifier qwen/qwen3.6-35b-a3b

These settings affect VRAM use, concurrency, context capacity, performance, and the stable model identifier used by existing applications. Falling back to LM Studio defaults after a PAIR or engine restart is not equivalent behavior.

Requested behavior

1. Allow load options to be configured per model and per node.
2. Persist those options across PAIR restarts, engine restarts, unload/reload operations, and machine reboots.
3. Apply the saved configuration when PAIR automatically loads or restores a model.
4. Display the effective load configuration and document which LM Studio options PAIR supports or intentionally ignores.

At minimum, support GPU offload, context length, parallelism, and a stable identifier. A generic advanced-arguments/configuration mechanism would also cover future lms load options.

Inizia con il manifest attuale del motore LM Studio e il relativo punto di ingresso lms load <model>. Traccia il modo in cui PAIR gestisce il download, il caricamento, l’espulsione, l’eliminazione e il ripristino automatico dei modelli, quindi determina dove possono persistere le impostazioni per nodo e per modello. Il lavoro è completato quando le opzioni di caricamento supportate sopravvivono agli scenari di riavvio e reboot elencati, vengono visualizzate efficacemente e sono documentate.