Stack from ghstack (oldest at bottom):

• -> [Inductor][CPP] Remove redundant Buffers after Grouped GEMM Fusion #143904
• [Inductor][CPP] Enable Epilogue Fusion for Grouped GEMM Template #143897
• [Inductor][CPP] Enable Grouped GEMM Template #143796

Summary
In this PR, we remove the extra kernel arguments and the extra buffers allocation when any MultiOutput Buffer is consumed by an out-template epilogue. If any MultiOutput Buffer is consumed by an out-template epilogue, the Grouped GEMM Template should bypass storing it in the MultiOutput Buffer and instead write it directly to the corresponding out-template epilogue.

Remove extra kernel arguments
For the case listed above, a MultiOutput Buffer shouldn't exist in the Kernel's args if it's consumed by an out-template epilogue. We mark this MultiOutput Buffer as REMOVED for this case.

Remove the extra buffers allocation
For the case listed above, a MultiOutput Buffer shouldn't be allocated. We introduce the outputs_removed attribute in the CppTemplateBuffer. This attribute tracks MultiOutput Buffers that are directly used by out-template epilogues. During code generation, if a MultiOutput Buffer is listed in outputs_removed, its buffer allocation line is omitted to prevent unnecessary memory usage.

Test Plan

python -u -m pytest -s -v test/inductor/test_cpu_select_algorithm.py -k test_grouped_linear_epilogue

cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @chauhang @aakhundov