v0.1.8 — multi-slot multi-agent, slot persistence, switch-mode drift fix

Re-tagged 2026-06-05 to include the wheel-build fix (drop redundant
hatch force-include for compat/profiles). The initial v0.1.8 tag was
deleted because its release-day build failed in GitHub Actions
(https://github.com/jaigouk/gpumod/actions/runs/27009703360).

- gpumod-8xaq: gemma4-26b-a4b-q4-multi preset (N=3 cont-batching, 128K per slot)
- gpumod-8viu: slot save/restore via --slot-save-path
- gpumod-hrgg: switch_mode reconciles DB vs running-set drift
- gpumod-kk2j: presets.md / architecture / cli / mcp-workflows refresh
- gpumod-hwbb: integration test for switch_mode partition
- gpumod-01lv: 8 new tutorials (getting-started + user-guide)
- CVE-clearing dep upgrade: authlib, fastmcp, pyjwt, urllib3, pyasn1, python-multipart
- Build fix: drop redundant force-include for compat/profiles

chore: bump version to 0.1.7

v0.1.6 - Multi-Agent Presets & CI Pipeline

New Features:
- Qwen3.5 multi-agent presets with 3 parallel slots:
  - qwen35-27b-multi (17GB VRAM, dense architecture)
  - qwen35-35b-multi (20GB VRAM, Q4_K_XL MoE)
  - qwen35-35b-q3-multi (16GB VRAM, Q3_K_XL MoE)
  - qwen3-coder-multi-p3 for code mode
- llama.cpp architecture compatibility checker with version detection
- Tetris multi-agent benchmark comparing 27B vs 35B models

Infrastructure:
- GitHub Actions CI with lint, typecheck, test, and audit gates
- PyPI publish workflow with trusted publishing

Fixes:
- MCP generate_preset now outputs flat PresetConfig format
- Code mode updated to use multi-slot preset

Benchmark Results:
- 35B MoE models are 3.8x faster than 27B dense for multi-agent tasks
- Q3 quantization shows minimal quality loss vs Q4 with 4GB VRAM savings

docs: add mixed workload benchmark results (2026-02-22)

chore: include uv.lock version bump to 0.1.4

v0.1.3: Multi-agent benchmark, unit file regeneration

Release v0.1.2

Highlights:
- Fixed orphan service cleanup during mode switches
- VRAM wait-for-release before starting new services
- Improved health check with per-service startup timeout
- SelfCheckGPT consistency-based hallucination detection
- Model compatibility pre-flight validation
- mkdocs exclude for QA session files
- Trivy misconfig scan fix