Refactor display device name / Fix tokenize_data from OpusMtEnESTransformer

[Irozuku]

This pull request refactors how device selection (CPU/GPU) is handled across several Hugging Face model classes in the backend. The main improvement is centralizing device enumeration and mapping logic into a new utility module, which simplifies device handling and ensures consistency. Additionally, the device selection in model initialization is now more robust, and the translation model's data tokenization logic is improved for clarity and correctness.

Device handling refactor:

• Introduced a new utility module utils.py that centralizes device enumeration, placeholder, and mapping logic (DEVICE_ENUM, DEVICE_PLACEHOLDER, NAME_TO_DEVICE) for CPU and available GPUs, using human-readable names for GPUs.
• Updated stable_diffusion_v1_depth_controlnet.py, stable_diffusion_v2_model.py, and stable_diffusion_v3_model.py to import device constants and mappings from the new utility module, removing redundant device code from each file. [1] [2] [3] [4]

Model initialization improvements:

• Modified model initializers to use the canonical device string from NAME_TO_DEVICE mapping, ensuring correct device assignment regardless of user input format. [1] [2] [3] [4]
• Adjusted device serialization and deserialization in model loading to use human-readable device names and proper mapping back to device strings.

Translation model tokenization fix:

• Refactored the tokenize_data method in OpusMtEnESTransformer to handle cases where target data (y) is missing more cleanly, and improved the logic for constructing tokenized samples.

Documentation update:

• Updated the device selection field description in OpusMtEnESTransformerSchema to clarify that selecting GPU will use all available GPUs.