update vllm_hpu_extension commit to 24039a3

[ccrhx4]

to fix part of the #443

update the vllm_hpu_extension commit id to HabanaAI/vllm-hpu-extension@24039a3

[madamczykhabana]

Revert for PR26 is pending: HabanaAI/vllm-hpu-extension#31

[michalkuligowski]

This needs to be rejected due to: HabanaAI/vllm-hpu-extension#31

[ccrhx4]

Inference accuracy is at the risk of being compromised when the softmax_mode == "fast".

From the Habana document, https://docs.habana.ai/en/latest/PyTorch/Model_Optimization_PyTorch/Optimization_in_PyTorch_Models.html#using-fused-sdpa

Using fast Softmax may affect inference accuracy.
Only BF16 data type is supported with fast Softmax.
Fast Softmax is not supported when running training in recompute mode with is_causal = False.

Even in the optimum-habana, the softmax_mode is default as "None". https://github.com/huggingface/optimum-habana/blob/f488ab66329b2c3a46063292b2822c1cab068756/optimum/habana/transformers/generation/configuration_utils.py#L38

Please reconsider this issue, I understand the performance is important, but the model needs to be performed correctly as it has been trained to.