Revert "add cache flush to flex att bs=16"

This reverts commit 19e276e.

Author: dev-tomek

benchmarks/triton_kernels_benchmark/flex_attention_benchmark_causal_mask.py

@@ -149,7 +149,6 @@ def causal_mask(_, __, q_idx, kv_idx):
     ))
 def benchmark(Z, H_q, H_kv, N_CTX_q, N_CTX_kv, D_HEAD_qk, D_HEAD_v, MODE, provider):
     # Maximum across torch=200, triton=600
-    torch.xpu.empty_cache()
     do_bench = benchmark_suite.get_do_bench(n_warmup=600, n_repeat=10, quantiles=[0.5, 0.0, 1.0])
     if MODE not in ('fwd', 'bwd'):
         raise ValueError(f"Invalid MODE: {MODE}. Expected 'fwd' or 'bwd'.")