feat:feat: Added logging, parallel processing, and CPU processing option for FP8 to BF16 conversion

[anand-144]

• Implemented logging to track the conversion process and handle missing tensors more effectively.
• Introduced parallel processing using ThreadPoolExecutor to speed up large model conversions.
• Added an option to use CPU (--use-cpu) instead of GPU for environments without CUDA support.
• Optimized memory usage by caching only the two most recently used files.
• Updated progress tracking with tqdm for better visibility.
• Ensured the model.safetensors.index.json file correctly removes _scale_inv references.
• These enhancements improve efficiency, usability, and robustness of the FP8 to BF16 conversion process.