feat: 🚀 allow per-call confidence_threshold override in prediction APIs

[onuralpszr]

AutoDetectionModel.from_pretrained() has always accepted confidence_threshold, but get_prediction() and get_sliced_prediction() had no way to override it without mutating the model object directly. This made dynamic threshold usage (e.g. sweeping thresholds, per-image tuning) awkward.

Both functions now accept an optional confidence_threshold parameter. When provided, it overrides detection_model.confidence_threshold for that call and restores the original value before returning, so the model state is unchanged after the call.

result = get_sliced_prediction(image, model, confidence_threshold=0.7)
# model.confidence_threshold is unchanged after this call

Related discussion topic: #1269

[onuralpszr]

PTAL @fcakyon when you had time.

sweep conf approach usage

from sahi import AutoDetectionModel
from sahi.predict import get_prediction
model = AutoDetectionModel.from_pretrained(
        model_type="ultralytics",
        model_path="yolo11n.pt",
        confidence_threshold=0.3,
    )
# run with a different threshold without changing the model
result = get_prediction("image.jpg", model, confidence_threshold=0.7)
print(model.confidence_threshold)  # still 0.3

[fcakyon]

looks good 👍🏻