[vLLM] add enforce eager as an option

[lanking520]

Description

This option will help to reduce CUDA graph GPU memory addition under some extreme scenarios

[maaquib]

Wondering if we should also add gpu-memory-utilization flag to configure more memory for CUDAGraph [link]

        logger.info("CUDA graphs can take additional 1~3 GiB memory per GPU. "
                    "If you are running out of memory, consider decreasing "
                    "`gpu_memory_utilization` or enforcing eager mode. "
                    "You can also reduce the `max_num_seqs` as needed "
                    "to decrease memory usage.")

[lanking520]

@maaquib there are negative impact on this parameter since the general pagedattention will need to use extra GPU spaces to get tokens calculated. I can add this, but not recommend to change on this for particular use cases, it's hard to balance the moving variants of two factors