Make acc matrix allocation on each call for XeTLA GEMM benchmarks

[anmyachev]

If we take for comparison: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/12382880184/job/34564504020 (main) vs https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/12390456716/job/34585505155 (PR), then the degradation from this pull request for XeTLA is ~3%.

However, this is also a potential opportunity to improve the Triton kernel by only allocating the accumulation matrix once. If this is implemented for Triton, this pull request will need to be rolled back for XeTLA.

CI runs:

• https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/12378323790 (upstream profiler CI)
• https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/12379634873
• https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/12381457034
• https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/12390411560 (upstream profiler) Wall time is used instead of elapsed_time (apparently chose the wrong runner)
• https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/12390456716 (legacy profiler)

[anmyachev]

Manual start, because for some reason the automatic one doesn't work: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/12392619525 (passed)