fix: handle empty inputs in OrientationPredictor

[badranX]

OrientationPredictor does not handle empty input batches and can raise an exception when orientation detection is enabled and no valid crops/images are passed to the predictor. This PR handles empty inputs similar to the RecognitionPredictor.

Reproduction

import requests

from doctr.io import DocumentFile
from doctr.models import ocr_predictor

receipt = requests.get(
    "https://github.com/mindee/doctr/releases/download/v0.3.0/mock_receipt.jpeg"
).content

predictor = ocr_predictor(
    pretrained=True,
    det_arch="fast_base",
    reco_arch="parseq",
    assume_straight_pages=False,
    detect_orientation=True,
    disable_crop_orientation=False,
    disable_page_orientation=False,
    straighten_pages=True,
)

docs = DocumentFile.from_images([receipt])
docs[0] = docs[0][:10, :10]

result = predictor(docs)

Result:

File "/home/ubuntu/dev/pr/doctr/doctr/models/preprocessor/pytorch.py", line 73, in batch_inputs
    if isinstance(samples[0], tuple):
                  ~~~~~~~^^^
IndexError: list index out of range