[LLM Runtime] Fix the accuracy issue with group_size=-1

[luoyu-intel]

Type of Change

update jblas' bug fix commits

[VincyZhang]

https://inteltf-jenk.sh.intel.com/view/nlp-toolkit-validation/job/ITREX-cpp-graph-extension/229/

[VincyZhang]

windows binary build: https://inteltf-jenk.sh.intel.com/view/nlp-toolkit-validation/job/nlp-release-build-win/299/

[DDEle]

It is confirmed that this PR benefits MLPerf accuracy:

• reference: "ROUGE1", 42.9865 * 0.99, "ROUGE2", 20.1235 * 0.99, "ROUGEL", 29.9881 * 0.99, "GEN_LEN", 4016878*0.9
• q4_j1_bf16_pc-gcc12 ([Cpp Graph] Align Cpp Beam Search  #322):
  {'rouge1': 43.3352, 'rouge2': 20.1831, 'rougeL': 29.8832, 'rougeLsum': 40.4381, 'gen_len': 4414947, 'gen_num': 13368}
• q4_j1_bf16_pc-no_mha (around PR950, 8b9020e) (all):
  {'rouge1': 42.8626, 'rouge2': 19.8439, 'rougeL': 29.4733, 'rougeLsum': 40.0532, 'gen_len': 4576100, 'gen_num': 13368}
  (less than 99%)
• q4_j_bf16_pc-no_mha-gcc13 (inside PR1057, 369c597) (all) ==>(falsely reported as j1 previously)<==
  {'rouge1': 43.2917, 'rouge2': 20.1802, 'rougeL': 29.9402, 'rougeLsum': 40.4317, 'gen_len': 4413826, 'gen_num': 13368}
• q4_j1_bf16_pc-no_mha-gcc13 (PR1059+, 6e3a514+)(all):
  {'rouge1': 42.8626, 'rouge2': 19.8439, 'rougeL': 29.4733, 'rougeLsum': 40.0532, 'gen_len': 4576100, 'gen_num': 13368}