feat(inference): add curated DGX Spark serving-profile choices

## Goal

Make the primary DGX Spark onboarding experience present these curated serving choices:

1. Muse
2. NVIDIA Nemotron 3.5 Lightning
3. DeepSeek V4 Flash

Keep the broader catalog discoverable and selectable through:

```bash
nemoclaw profiles list
nemoclaw onboard --profile
```

## Scope

- Add a Spark-specific curated profile picker to interactive onboarding without changing non-interactive defaults.
- Map Muse and Lightning to single-DGX-Spark serving profiles.
- Map DeepSeek to the existing two-DGX-Spark managed-cluster profile, and make its topology requirement explicit in the picker.
- Preserve `--profile` as the stable automation and advanced-selection path.
- Add unit coverage for ordering, selection, invalid input, non-interactive bypass, explicit-profile bypass, and required topology handling.
- Document support states and hardware requirements.

## DeepSeek validation boundary

The upstream vLLM quickstart documents DeepSeek V4 Flash on 4x B200/B300. NemoClaw's existing Experimental two-DGX-Spark `DeepSeek-V4-Flash-0731` DSpark recipe is therefore the candidate for validation; do not expose DeepSeek as a one-DGX-Spark profile without qualification.

## Acceptance criteria

- Interactive DGX Spark onboarding shows Muse, Lightning, and DeepSeek in the order above.
- DeepSeek is unavailable or clearly rejected unless the required paired-Spark topology is configured and qualified.
- `nemoclaw profiles list` continues to list all installed profiles with compatibility results.
- `nemoclaw onboard --profile ` remains the documented explicit path.
- Existing default and non-interactive onboarding behavior remains unchanged.