feat: add DeepSeek V4 Flash dual-Spark deployment

[xu16601526267]

Summary

• add catalog assets for deepseek-v4-flash-0731, the Anemll DSpark GB10 engine, and a worker-first two-DGX-Spark deployment scenario
• extend deployment scenarios with typed inputs, unique deployment IDs, per-deployment environment overrides, and local/remote Fleet targets
• expose scenario-backed models in the deploy modal so the worker and RoCE settings can be reviewed, dry-run, saved locally, and launched from the UI
• preserve integral JSON/YAML config values such as max_model_len=1048576 as decimal CLI arguments instead of scientific notation
• let engine YAML declare container-specific host networking, devices, shared memory, and ulimits; render quoted command arguments safely in K3S pods

Scope

The catalog contains no machine-specific usernames, IP addresses, paths, benchmark artifacts, or captured request/response evidence. Fabric addresses, interfaces, HCAs, GID indexes, and the Fleet worker ID are scenario inputs.

The two-node scenario intentionally uses no_pull: true: the image and checkpoint must be available on both devices before worker-first startup, avoiding a rendezvous timeout while the head downloads a large checkpoint.

Validation

• go vet ./...
• focused tests for catalog loading/resolution, decimal flag rendering, Docker/K3S container options, Fleet scenario routing, MCP/CLI wiring, model annotation, and deploy UI rendering
• JavaScript syntax check for the embedded UI

go test ./... was also attempted on Windows; unrelated existing platform-specific tests fail there on symlink privileges, Unix sleep, port/process assumptions, and temporary SQLite cleanup. The packages and cases changed by this PR pass.