feat(ml): GPU-capable tagger + embedder with CPU fallback (#872)

Step 1 of GPU enablement (code only — CPU-safe, CI-green; the CUDA image is a
separate step pending the host driver version).

- New services/ml/device.py: FC_ML_DEVICE (auto|cuda|cpu) intent + VRAM knobs
  (FC_ML_ONNX_GPU_MEM_GB, FC_ML_TORCH_MEM_FRACTION). Per-worker-host bootstrap →
  env, not a DB setting (the GPU host runs CUDA, others CPU).
- tagger: use CUDAExecutionProvider (with gpu_mem_limit) when requested AND the
  provider is actually present (onnxruntime-gpu), else CPUExecutionProvider. Logs
  the active providers.
- embedder: move model + inputs to cuda when requested AND torch.cuda is
  available; cap torch's VRAM share; .detach().cpu() before numpy. fp32 kept so
  GPU embeddings stay in the same space as existing CPU ones.

Both AND the env intent with the framework's real availability, so on CPU
(CI / CPU onnxruntime / no GPU) they fall back cleanly — behavior unchanged.
The 8GB P4 is shared by both frameworks, hence the conservative default caps.

Tests: device env parsing. (tagger/embedder GPU paths are operator-verified on
the GPU host — models aren't in CI.)

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

backend/app/services/ml/device.py

@@ -0,0 +1,31 @@
+"""ML device selection (#872 — GPU enablement for the ml-worker).
+
+The ml-worker is GPU-capable but must run unchanged on CPU (CI, non-GPU hosts).
+Selection is a per-worker-HOST bootstrap concern (the GPU host runs CUDA, others
+CPU), so it's an env var, not a DB setting — different workers need different
+values. Each framework still ANDs this intent with its OWN runtime availability
+(onnxruntime providers / torch.cuda), so "want GPU but none present" falls back
+to CPU cleanly.
+
+Env:
+  FC_ML_DEVICE            auto (default) | cuda | gpu  -> try GPU;  cpu -> force CPU
+  FC_ML_ONNX_GPU_MEM_GB   ONNX CUDA arena cap, GB (default 3) — the P4 is 8GB
+                          total and torch shares it, so keep headroom.
+  FC_ML_TORCH_MEM_FRACTION fraction of total VRAM torch may use (default 0.6).
+"""
+
+import os
+
+
+def gpu_requested() -> bool:
+    return os.environ.get("FC_ML_DEVICE", "auto").strip().lower() in (
+        "auto", "cuda", "gpu",
+    )
+
+
+def onnx_gpu_mem_bytes() -> int:
+    return int(float(os.environ.get("FC_ML_ONNX_GPU_MEM_GB", "3")) * 1024 ** 3)
+
+
+def torch_mem_fraction() -> float:
+    return float(os.environ.get("FC_ML_TORCH_MEM_FRACTION", "0.6"))