Add LaMa magic eraser, remote provider abstraction, and AI tool infrastructure

Backend:
- requirements.txt: add simple-lama-inpainting, rembg[gpu]; upgrade opencv to 4.10+
- app/config.py: add InvokeAI (url, model) and ComfyUI (url, model) settings; OPENAI_MODEL
- app/services/local_inpaint.py: LaMa, OpenCV, rembg wrappers (auto GPU/CPU)
- app/services/remote_provider.py: abstract RemoteAIProvider + OpenAI, InvokeAI, ComfyUI drivers
- app/routers/ai_tools.py: new /api/* endpoints — /erase, /inpaint/lama, /inpaint/fast,
  /background/remove, /inpaint/remote, /generate/txt2img, /generate/img2img,
  /generate/outpaint, GET /config (capability flags)
- app/main.py: register ai_tools router

Frontend:
- services/api.js: add erase(), textToImage(), imageToImage(), remoteInpaint(), getConfig()
- api/capabilities.js: lazy-fetch /api/config singleton; hasRemote() helper
- tools/ai_lama_erase.js: brush-paint mask → LaMa erase → apply to layer
- tools/ai_smart_inpaint.js: brush mask + dialog (Fast/Quality mode + prompt) → inpaint
- core/components/provider-badge.js: shows active provider + health in toolbar
- config.js: register ai_lama_erase and ai_smart_inpaint tools
- main.js: mount provider badge on load
- .env.example: document InvokeAI, ComfyUI, OpenAI provider settings

https://claude.ai/code/session_01B58MaJCU1R6KwBDJCp8AfN

backend/app/services/local_inpaint.py

@@ -0,0 +1,82 @@
+"""
+Local inpainting operations — LaMa, OpenCV, and background removal.
+All operations use GPU automatically if PyTorch detects one, CPU otherwise.
+"""
+
+from io import BytesIO
+from PIL import Image
+import numpy as np
+import cv2
+
+# Lazy-loaded LaMa model (downloaded on first use, ~100MB)
+_lama = None
+
+
+def get_lama():
+    global _lama
+    if _lama is None:
+        from simple_lama_inpainting import SimpleLama
+        _lama = SimpleLama()
+    return _lama
+
+
+def lama_available() -> bool:
+    try:
+        import simple_lama_inpainting  # noqa: F401
+        return True
+    except ImportError:
+        return False
+
+
+def lama_inpaint(image_bytes: bytes, mask_bytes: bytes) -> bytes:
+    """LaMa structural inpainting — best for object removal and large fills."""
+    lama = get_lama()
+    image = Image.open(BytesIO(image_bytes)).convert("RGB")
+    mask = Image.open(BytesIO(mask_bytes)).convert("L")
+    if mask.size != image.size:
+        mask = mask.resize(image.size, Image.Resampling.LANCZOS)
+    result = lama(image, mask)
+    buf = BytesIO()
+    result.save(buf, format="PNG")
+    return buf.getvalue()
+
+
+def opencv_inpaint(image_bytes: bytes, mask_bytes: bytes, method: str = "telea") -> bytes:
+    """OpenCV fast structural inpainting — CPU only, milliseconds."""
+    image = Image.open(BytesIO(image_bytes)).convert("RGB")
+    mask = Image.open(BytesIO(mask_bytes)).convert("L")
+    if mask.size != image.size:
+        mask = mask.resize(image.size, Image.Resampling.LANCZOS)
+
+    img_np = np.array(image)
+    mask_np = np.array(mask)
+    _, mask_bin = cv2.threshold(mask_np, 127, 255, cv2.THRESH_BINARY)
+
+    flags = cv2.INPAINT_TELEA if method == "telea" else cv2.INPAINT_NS
+    result = cv2.inpaint(img_np, mask_bin, inpaintRadius=3, flags=flags)
+
+    buf = BytesIO()
+    Image.fromarray(result).save(buf, format="PNG")
+    return buf.getvalue()
+
+
+def remove_background_rembg(image_bytes: bytes) -> bytes:
+    """Background removal using rembg."""
+    from rembg import remove
+    return remove(image_bytes)
+
+
+def rembg_available() -> bool:
+    try:
+        import rembg  # noqa: F401
+        return True
+    except ImportError:
+        return False
+
+
+def gpu_available() -> bool:
+    try:
+        import torch
+        return torch.cuda.is_available()
+    except ImportError:
+        return False