v0.1.0

The first public release.

Added

• A dashboard for llama.cpp servers: per-request decode and prefill speed,
  latency, context depth, speculative-decoding acceptance, slots, the KV cache
  and the loaded model's identity and shape.
• GPU telemetry for NVIDIA cards against each card's own limits, host CPU,
  memory, disks and network, and BMC sensors and event log where there is one.
• Alerts that hold back during a restart or model switch, and a learned
  per-model speed baseline that catches a silent fallback to the CPU.
• Push notifications through Apprise URLs (NOTIFY_URLS) or a command.
• History for up to two years in SQLite, with 5-minute and hourly rollups.
• An agent that finds llama.cpp in a Docker container, a systemd unit or a
  Proxmox LXC container, and follows its log for per-request timings.
• A dashboard-only mode against any llama-server, with no agent.
• compose.yaml for one Docker host; deploy/install-agent.sh for the agent
  as a systemd service; deploy/proxmox/install.sh for Proxmox.
• Container images for amd64 and arm64 on GHCR.
• Demo mode (DEMO=1), which replays a recorded session with no GPU needed.
• /metrics in Prometheus format.