feat(catalog): add GLM-5.3 Flash dual-Spark deployment

[xu16601526267]

Summary

• add the glm-5.3-flash-nvfp4 model asset and bind it to a two-device deployment scenario
• add the validated GB10 vLLM engine defaults: TP2, 256K context, FP8 KV, MTP-4, 128-token prefix matching, prompt-cache telemetry, multimodal limits, and GLM tool/reasoning parsers
• add a worker-first AIMA Fleet scenario with typed RoCE address, interface, HCA, and GID inputs
• verify catalog resolution and generated container arguments, including the 128 / index_kpool=4 boundary and checkpoint-declared NVFP4 handling

Scope

The catalog contains no machine-specific usernames, IP addresses, paths, credentials, benchmark artifacts, logs, or captured requests.

The engine image is intentionally declared as a local distribution and both scenario deployments use no_pull: true. The validated aima/glm53-flash:sm121-v8-local cache-safety image and the checkpoint must be installed on both Fleet devices before worker-first startup. This avoids claiming that the older public v8 image contains the newer prefix-cache and block-table safety fixes.

Validation

• go test ./internal/knowledge -count=1
• go test ./cmd/aima -run 'Test(AnnotateModelsFromCatalogAddsDeploymentScenario|Scenario.*)$' -count=1
• go vet ./...
• git diff --check