Align LLM.generate with Modern vLLM API for Static mode

[WarningRan]

Summary

Update the inference driver to resolve TypeError caused by the deprecation of the prompt_token_ids keyword argument in recent vLLM versions.

Changes

• Input Refactoring: Switched from passing raw token lists via prompt_token_ids=... to the modern prompts parameter using the PromptDict format: [{"prompt_token_ids": [...]}, ...].
• Generator Update: Enhanced RandomGenerator to directly yield vLLM-compliant data structures, reducing overhead during benchmarking.
• API Alignment: Aligned LLM.generate calls with the Sequence[PromptType] specification to ensure future-proof compatibility.

Comparison

Feature	Before (Legacy)	After (Modern)
Argument	prompt_token_ids=inputs	prompts=[{"prompt_token_ids": i} for i in inputs]
Data Type	NumPy Array / Raw List	List of Dictionaries (PromptType)
Stability	Throws TypeError in new vLLM	Fully compatible with 0.6.0+

Notes

• Tested Environment: This change has been verified on the Spyre vLLM image.
• Compatibility: While optimized for Spyre, the logic follows standard vLLM protocols and is expected to be portable to other backends.

GitHub Issue(s) to reference

Reminders

• should this PR noted in the Changelog?