immich: GPU-accelerate video transcoding (NVENC + NVDEC)

Pin immich-server to the GPU node with a time-sliced nvidia.com/gpu slice
so ffmpeg uses hardware NVENC encode + NVDEC decode instead of software.
This frees the ~3-4 CPU cores the software transcoder was burning inside
the request-serving pod (which was slowing thumbnail/photo browsing), and
makes incompatible (HEVC/iPhone) videos playable in seconds. Activation is
ffmpeg.accel=nvenc + accelDecode=true in the DB system-config (Immich app
config is DB-managed here, like oauth/smtp — not Terraform).

Also give immich-frame the same Keel ignore_changes immich-server already
has, so an untargeted apply no longer churns it (pre-existing drift).

Docs: .claude/CLAUDE.md Immich row + compute.md GPU-workloads list.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

docs/architecture/compute.md

@@ -330,10 +330,14 @@ label with it, and `null_resource.gpu_node_config` re-applies the
 next apply (discovery keyed on
 `feature.node.kubernetes.io/pci-10de.present=true`).
 
-**GPU Workloads**:
-- Ollama (LLM inference)
-- ComfyUI (Stable Diffusion workflows)
-- Stable Diffusion WebUI
+**GPU Workloads** (time-sliced — node advertises `Tesla-T4-SHARED`,
+`sharing-strategy=time-slicing`, `nvidia.com/gpu.replicas=100`, so many pods
+share the single T4; request `nvidia.com/gpu: 1` for a slice, not the whole card):
+- immich-machine-learning (CLIP smart-search + facial recognition, CUDA)
+- immich-server (NVENC/NVDEC video transcoding — `ffmpeg.accel=nvenc` + `accelDecode=true`)
+- Frigate (object-detection inference)
+- llama-cpp / llama-swap (LLM inference)
+- nvidia-exporter + gpu-pod-exporter (DCGM metrics)
 
 ## Configuration