make both TP overlap splitting by batch and seq work together

[xuechendi]

VLLM_TP_SPLIT_SIZE_BY_SEQ=2
VLLM_TP_SPLIT_SIZE_BY_BATCH=2

[xuechendi]

@jikunshang , please take a review.
when both set 2, we will only pick one strategy: I found use both hurt the performance
Will try split_by batch firstly, since it improves performance more
if split_by_batch is performed, hint to disable split_by_seq; otherwise split by sequence

[xuechendi]

One lacking issue, this PR revert Prepare_cos_sin => which shows slightly performance regression, need to fix that before merge to mlperf branch

[jikunshang]

LGTM.