Fixes a crash when some labels/boxes are NaN

[BloodAxe]

Test dataset to reproduce. See the attached report generated Report.pdf

If batch/sample with NaN is detected it is excluded from processing entirely. But it is tracked and added to list of errors which is not also written to PDF

import numpy as np
import torch

from data_gradients import DetectionAnalysisManager

train_samples = [
    # 4 images with labels
    (torch.randn((3, 100, 100), dtype=torch.float32), torch.randint(0, 10, (1, 5), dtype=torch.float32)),
    (torch.randn((3, 100, 100), dtype=torch.float32), torch.randint(0, 10, (2, 5), dtype=torch.float32)),
    (torch.randn((3, 100, 100), dtype=torch.float32), torch.randint(0, 10, (3, 5),

 dtype=torch.float32)),
    (torch.randn((3, 100, 100), dtype=torch.float32), torch.randint(0, 10, (4, 5), dtype=torch.float32)),
]
train_samples[3][1][0:2, 2] = np.nan

valid_samples = [
    # 4 images with labels
    (torch.randn((3, 100, 100), dtype=torch.float32), torch.randint(0, 10, (1, 5), dtype=torch.float32)),
    (torch.randn((3, 100, 100), dtype=torch.float32), torch.randint(0, 10, (2, 5), dtype=torch.float32)),
    (torch.randn((3, 100, 100), dtype=torch.float32), torch.randint(0, 10, (3, 5), dtype=torch.float32)),
    (torch.randn((3, 100, 100), dtype=torch.float32), torch.randint(0, 10, (4, 5), dtype=torch.float32)),
]
valid_samples[2][1][0, 2:4] = np.nan
manager = DetectionAnalysisManager(
    train_data=train_samples,
    val_data=valid_samples,
    report_title="Detection Test",
    class_names=["class_1", "class_2", "class_3", "class_4", "class_5", "class_6", "class_7", "class_8", "class_9",
                 "class_10"],
    batches_early_stop=None,
)
manager.run()