[Enhancement] aglin online quantizaiton and offline quantization

[haoyangli0109]

The main difference in precision stems from the different round_mode settings used during scale calculations. Previously, we used Aiter’s mxfp4 kernel for online quantization, which uses round_up, whereas the offline Quark kernel uses round_even. In terms of precision, round_even is more advantageous.
This PR fixes this issue, and switching the kernel does not result in any additional overhead for online quantization.

Please note that you need to use gsm8kshot3 to reproduce the issue; gsm8kshot5 has nearly identical accuracy.

python3 -m atom.entrypoints.openai_server --model /shareddata/deepseek-ai/DeepSeek-R1-0528 \
  --enforce-eager -tp 8 \
  --port 5679 --server-port 7778 \
  --online_quant_config '{"global_quant_config":"ptpc_fp8","layer_quant_config":{"*expert*":"mxfp4"},"exclude_layer":["lm_head","*.gate.*"]}' \
  --method mtp --num-speculative-tokens 3 


lm_eval \
  --model local-completions \
  --model_args "model=/shareddata/deepseek-ai/DeepSeek-R1-0528,base_url=http://localhost:7778/v1/completions,tokenized_requests=False,tokenizer_backend=None,num_concurrent=32" \
  --tasks gsm8k \
  --num_fewshot 3 \
  --batch_size auto

before

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     3|exact_match|↑  |0.9333|±  |0.0069|
|     |       |strict-match    |     3|exact_match|↑  |0.9287|±  |0.0071|

modified:

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     3|exact_match|↑  |0.9416|±  |0.0065|
|     |       |strict-match    |     3|exact_match|↑  |0.9401|±  |0.0065|