GPU selection specific to embedding models

[thakurvivek]

Is your pull request related to a problem? Please describe.
No ability to choose GPU while loading embedding models

Why should this feature be added?
In multi-GPU environments, TabbyAPI currently lacks the ability to specify which GPU device should be used for embedding models. This creates significant resource management challenges:

• Current Limitation: Embedding models automatically load on default GPU (typically GPU 0)
• Resource Conflict: Cannot distribute embedding workloads across available GPUs
• Operational Inefficiency: Forces users to manually manage GPU allocation via environment variables

Examples
In a multi-GPU cluster setup:

• GPU 3090: Running large language models (high memory usage)
• GPU 5090: Dedicated for embedding models (currently underutilized)
• Current TabbyAPI instance: Cannot target GPU 5090 for embeddings without restarting

Additional context
Added embeddings_device_id configuration parameter to enable specific GPU selection