refactor(services): consolidate model-serving projects under services

[wenxind-nvidia]

Summary

• move the VLM, STT, three LLM, and two TTS projects into direct-child services/ directories
• update model-servers, Simple VLM, launcher references, presets, copied YAML comments, GPU tests, CI/tooling paths, notices, and dependency documentation
• preserve package names, commands, ports, and serving behavior
• normalize standalone reference YAMLs to the repository-root models/ cache and document how existing offline installations can merge ignored caches left under ai-services/
• document and test the direct-child service convention, including a tracked-text guard against retired service paths

Cache migration

Git cannot move ignored model weights. Before starting relocated services offline, users of the old standalone VLM, STT, Nemotron 3 Nano, Nemotron Omni, Magpie TTS, or Piper TTS configurations should merge ai-services/models/ and ai-services/tts/models/ into models/. The exact non-clobbering procedure is documented in docs/ai-services.md and linked from services/README.md.

Sample-owned YAMLs and the standalone Llama Nemotron YAML already resolved to repository-root models/ and need no migration.

Scope

• MCP adapters remain under agent-mcp-servers/
• xr-render-demo changes are limited to mechanical model-service path replacements in main.py and one stale path correction in a YAML comment
• no render worker, scene, prompt, eval, or architecture changes

Validation

• uv run --project tests pytest -q tests -m 'not gpu' — 692 passed, 2 skipped, 10 deselected
• focused service-layout, model-config, and launcher tests — 53 passed
• uv run pytest -q test_piper_tts.py — 1 passed
• uv lock --project for all seven moved projects
• uvx ruff check --no-cache .
• strict Sphinx build with warnings as errors
• SPDX header check and git diff --check

Stacked on PR #320.