docs: document convert command and GPU pre-convert workflow

[michaelalber]

Summary

• Adds convert command to architecture diagram as an optional GPU pre-processing step
• Documents all six CLI flags (--collection, --force, --dry-run, --no-ocr, path arg)
• Explains sidecar workflow: convert once on GPU; ingest reads sidecars and skips Docling
• Documents optional flash-attn / cuda_use_flash_attention2 acceleration for Ampere+ GPUs
• Adds two troubleshooting rows: per-file crash isolation and slow conversion without flash-attn
• Updates architecture prose from two-process to three-process description
• Adds two feature bullets: GPU-accelerated pre-convert and crash-isolated batch conversion

Test plan

• Verify rendered Markdown looks correct on GitHub (architecture diagram ASCII art, code blocks)
• Confirm all convert CLI flags match __main__.py implementation

🤖 Generated with Claude Code