cpu: aarch64: enable bf16f32 matmul

benchdnn: `ONEDNN_VERBOSE=all ./benchdnn --matmul --dt=bf16:bf16:f32 32x128:128x32`
Before:
 onednn_verbose,v1,primitive,create:dispatch,matmul,cpu,matmul,gemm:acl,undef,src:bf16:a:any:any::f0 wei:bf16:a:any:any::f0 dst:f32:a:any:any::f0,,,32x128:128x32,unsupported datatype combination,src/cpu/aarch64/matmul/acl_matmul.cpp:89
After: (picks up ACL kernel)
 onednn_verbose,v1,primitive,exec:external,CpuGemmAssemblyWrapperKernel/sve_ffhybrid_bf16fp32_mmla_6x4VL,0.242188
 onednn_verbose,v1,primitive,exec,cpu,matmul,gemm:acl,undef,src:bf16:a:blocked:ab::f0 wei:bf16:a:blocked:BA8b4a::f0 dst:f32:a:blocked:ab::f0,,,32x128:128x32,0.267822

src/cpu/aarch64/matmul/acl_matmul.cpp

@@ -76,12 +76,18 @@ status_t acl_matmul_t::pd_t::init(engine_t *engine) {
             = utils::everyone_is(data_type::bf16, src_md()->data_type,
                       weights_md()->data_type, dst_md()->data_type)
             && platform::has_data_type_support(data_type::bf16);
+    const bool is_bf16f32_ok
+            = utils::everyone_is(data_type::bf16, src_md()->data_type,
+                      weights_md()->data_type)
+            && utils::everyone_is(data_type::f32, dst_md()->data_type)
+            && platform::has_data_type_support(data_type::bf16);
 
     // we need to save this state as it can change inside set_default_formats()
     weights_format_kind_ = weights_md_.format_kind;
 
     VDISPATCH_MATMUL(is_dense_format_kind(), VERBOSE_UNSUPPORTED_SPARSE_CFG);
-    VDISPATCH_MATMUL(utils::one_of(true, is_fp32_ok, is_fp16_ok, is_bf16_ok),
+    VDISPATCH_MATMUL(utils::one_of(true, is_fp32_ok, is_fp16_ok, is_bf16_ok,
+                             is_bf16f32_ok),
             VERBOSE_UNSUPPORTED_DT_CFG);
     VDISPATCH_MATMUL(!has_zero_dim_memory(), VERBOSE_EMPTY_TENSOR, "");
     VDISPATCH_MATMUL(set_default_formats(), VERBOSE_UNSUPPORTED_TAG);