Enhance GPU Management, Operator Control, and Multiprocessing

[drcege]

This PR enhances GPU resource management, individualized operator control, and robust multiprocessing support.

Optimize EasyOCR GPU Memory Usage

• Initialize EasyOCR on the CPU first, then explicitly set the detector and device to the target GPU. This avoids default data parallelism across all GPUs.
• Closes video_ocr_area_ratio_filter may cause cuda out of memory #353.

Fix Per-Operator Accelerator Support

• Remove global use_cuda() checks and switch to per-operator checks within the executor loop.
• Implement a per-operator op.use_cuda() method in the base OP class, considering both GPU availability and self.accelerator attribute.
• Refactor accelerator parsing and configuration logic to ensure proper setup, and ensure compatibility with jsonargparse for argument registration.
• Add missing accelerator configurations for CUDA-related operators. @HYLcool, please verify if any are still missing.

Implement Per-Operator Multiprocessing Method

• Introduce a UNFORKABLE decorator to record operators incompatible with the fork mode.
• Shift from a global multiprocessing strategy to per-operator runtime settings.
• Update unit tests to automatically switch to spawn mode based on GPU availability.

[yxdyc]

LGTM