fix: Return 400 instead of 500 for raw bytes sent as base64 image

[bigbitbus]

Summary

• When a client sends raw image bytes (e.g. JPEG starting with 0xFF 0xD8) with type: "base64", load_image_base64 calls value.decode("utf-8") which throws an unhandled UnicodeDecodeError
• This falls through to the catch-all exception handler and returns a 500 Internal Error with no useful message
• This fix catches UnicodeDecodeError and raises InputImageLoadError, which the existing error handler returns as a 400 with a clear message telling the client to base64-encode the image

Example stack trace we are fixing

\xff\\xd9', 0, 1, 'invalid start byte')"], "event": "%s: %s", "request_id": "18e5e45cade4464e9ca3d81481667ac4", "timestamp": "2026-02-18 09:45.08", "exception": {"type": "UnicodeDecodeError", "message": "'utf-8' codec can't decode byte 0xff in position 0: invalid start byte", "stacktrace": [{"filename": "/var/task/inference/core/interfaces/http/error_handlers.py", "lineno": 83, "function": "wrapped_route", "code": "return route(*args, **kwargs)"}, {"filename": "/var/task/inference/usage_tracking/collector.py", "lineno": 715, "function": "sync_wrapper", "code": "res = func(*args, **kwargs)"}, {"filename": "/var/task/inference/core/interfaces/http/http_api.py", "lineno": 3136, "function": "legacy_infer_from_request", "code": "inference_response = self.model_manager.infer_from_request_sync("}, {"filename": "/var/task/inference/core/managers/decorators/fixed_size_cache.py", "lineno": 177, "function": "infer_from_request_sync", "code": "return super().infer_from_request_sync(model_id, request, **kwargs)"}, {"filename": "/var/task/inference/core/managers/decorators/base.py", "lineno": 106, "function": "infer_from_request_sync", "code": "return self.model_manager.infer_from_request_sync(model_id, request, **kwargs)"}, {"filename": "/var/task/inference/core/managers/active_learning.py", "lineno": 54, "function": "infer_from_request_sync", "code": "prediction = super().infer_from_request_sync("}, {"filename": "/var/task/inference/core/managers/base.py", "lineno": 265, "function": "infer_from_request_sync", "code": "rtn_val = self.model_infer_sync("}, {"filename": "/var/task/inference/core/managers/base.py", "lineno": 338, "function": "model_infer_sync", "code": "return model.infer_from_request(request)"}, {"filename": "/var/task/inference/core/models/base.py", "lineno": 134, "function": "infer_from_request", "code": "responses = self.infer(**request.dict(), return_image_dims=False)"}, {"filename": "/var/task/inference/core/models/instance_segmentation_base.py", "lineno": 97, "function": "infer", "code": "return super().infer("}, {"filename": "/var/task/inference/core/models/roboflow.py", "lineno": 821, "function": "infer", "code": "return super().infer(image, **kwargs)"}, {"filename": "/var/task/inference/usage_tracking/collector.py", "lineno": 715, "function": "sync_wrapper", "code": "res = func(*args, **kwargs)"}, {"filename": "/var/task/inference/core/models/base.py", "lineno": 25, "function": "infer", "code": "preproc_image, returned_metadata = self.preprocess(image, **kwargs)"}, {"filename": "/var/task/inference/core/models/instance_segmentation_base.py", "lineno": 200, "function": "preprocess", "code": "img_in, img_dims = self.load_image("}, {"filename": "/var/task/inference/core/models/roboflow.py", "lineno": 1042, "function": "load_image", "code": "img_in, img_dims = self.preproc_image("}, {"filename": "/var/task/inference/core/models/roboflow.py", "lineno": 514, "function": "preproc_image", "code": "np_image, is_bgr = load_image("}, {"filename": "/var/task/inference/core/utils/image_utils.py", "lineno": 93, "function": "load_image", "code": "np_image, is_bgr = load_image_with_known_type("}, {"filename": "/var/task/inference/core/utils/image_utils.py", "lineno": 176, "function": "load_image_with_known_type", "code": "image = loader(value, cv_imread_flags)"}, {"filename": "/var/task/inference/core/utils/image_utils.py", "lineno": 271, "function": "load_image_base64", "code": "value = value.decode(\"utf-8\")"}]}, "filename": "error_handlers.py", "func_name": "wrapped_route", "lineno": 382}```

## Test plan
- [ ] Send raw JPEG bytes with `type: "base64"` to a legacy inference endpoint — should now return 400 with descriptive error instead of 500
- [ ] Send a valid base64-encoded image — should continue to work as before
- [ ] Send a malformed base64 string — should still return the existing 400 error from the `binascii.Error` handler

🤖 Generated with [Claude Code](https://claude.com/claude-code)