Backport cpu: aarch64: matmul: Optimize (A^T)*(B^T) and fuse sum post-op in acl_matmul

[fadara01]

Description

This is a backport of #1889 and #1892

Backport cpu: aarch64: matmul: Optimize (A^T)*(B^T) in acl_matmul
Computes (BA)^T instead of (A^T)(B^T) when the cost of transposing (B*A) is cheaper. This improves performance by ~1.25x for square matrices and even higher for
tall-skinny/fat-short matrices.

It also reduces code duplication and moves allocation of dst accumulator from ACL to scratchpad memory in oneDNN

Backport cpu: aarch64: matmul: fuse sum post op in acl matmul
Fuse the sum post op in acl matmul by setting the accumulate flag to
true in arm_compute::GEMMInfo. This speeds up the post op and saves
allocating a temporary dst sized tensor.

We also added _for_sum to use_dst_acc flag to stop it being confused
with the dst_acc used for transposing.

Change the way we deal with fused eltwise (as well as the new sum) to
fix segfaults when binary ops followed fused ops.

Fixes # (github issue)

Checklist

General

• [ YES ] Do all unit and benchdnn tests (make test and make test_benchdnn_*) pass locally for each commit?
• [ YES ] Have you formatted the code using clang-format?

Performance improvements

• [ YES ] Have you submitted performance data that demonstrates performance improvements?

New features

• Have you published an RFC for the new feature?
• Was the RFC approved?
• Have you added relevant tests?

Bug fixes

• Have you included information on how to reproduce the issue (either in a github issue or in this PR)?
• Have you added relevant regression tests?

RFC PR

• Does RFC document follow the template?
• Have you added a link to the rendered document?