grafana: replace custom metric collectors with community exporters

Replace three custom Prometheus textfile collector scripts with
dedicated community-maintained exporters:

- jellyfin-collector.nix (25 LoC shell) -> rebelcore/jellyfin_exporter
  Metric: jellyfin_active_streams -> count(jellyfin_now_playing_state)
  Bonus: per-session labels (user, title, device, codec info)

- qbittorrent-collector.nix (40 LoC shell) -> anriha/qbittorrent-metrics-exporter
  Metric: qbittorrent_{download,upload}_bytes_per_second -> qbit_{dl,up}speed
  Bonus: per-torrent metrics with category/tag aggregation

- intel-gpu-collector.nix + .py (130 LoC Python) -> mike1808/igpu-exporter
  Metric: intel_gpu_engine_busy_percent -> igpu_engines_busy_percent
  Bonus: persistent daemon vs oneshot timer, no streaming JSON parser

All three run as persistent daemons scraped by Prometheus, replacing
the textfile-collector pattern of systemd timers writing .prom files.
Dashboard PromQL queries updated to match new metric names.

services/grafana/default.nix

@@ -3,10 +3,8 @@
     ./grafana.nix
     ./prometheus.nix
     ./dashboard.nix
-    ./jellyfin-collector.nix
+    ./exporters.nix
     ./jellyfin-annotations.nix
-    ./qbittorrent-collector.nix
-    ./intel-gpu-collector.nix
     ./disk-usage-collector.nix
     ./llama-cpp-annotations.nix
     ./zfs-scrub-annotations.nix