fix: skip probing for incompatible scenario

[zhuangqh]

When determining the max available context length, vllm will
generate dummy inputs with max_num_batched_tokens tokens
and profiling peak VRAM usage.

1. skip probing for multi-gpu serving

If tensor-parallelism is enabled, there are more than 1 gpu workers.
However, we have no way to adjust the context length for profiling
for these workers in vllm.

2. skip probing when chunked-prefill is enabled

chunked prefill is a feature
to improve thoughput for long context serving (max_model_len > 32K).
It makes the peak memory usage nonlinearly related to max_model_len.

Notes for Reviewers:

test case:

• single gpu with enough vram to run max-model-len
• single gpu, find max-model-len after several probing
• single gpu, unable to run with even small model len
• multi gpu
• chunked prefill enabled