Allow users to configure vLLM runtime parameters

[anfredette]

vLLM has many configuration knobs (e.g., max_num_seqs, gpu_memory_utilization, enable_prefix_caching) that impact performance. We should expose relevant parameters for user customization.

Acceptance Criteria

• Identify most impactful vLLM configuration parameters
• Add fields to DeploymentIntent schema for vLLM config
• Update KServe InferenceService template to include vLLM args
• Provide sensible defaults based on use case and traffic profile
• Add UI controls for adjusting vLLM parameters (advanced mode)
• Document each parameter and its impact on performance

Notes

• Start with most critical parameters (e.g., KV cache settings, memory utilization)
• Advanced users may want fine-grained control
• Consider auto-tuning parameters based on traffic profile in future