cpu: x64: limit use of scratchpad in amx conv kernels

To avoid excessive memory consumption, limit size of scratchpad
to that of available L2 + L3 cache.

gpu: ocl: gen9 conv - fix performance issue

* prevent compiler to create redundant code when post ops are disabled

src: cpu: aarch64: add checks for supported int8 ACL activations

cpu: x64: jit_avx2: fix segfault for binary post-ops with nspc layout

gpu, binary: fix for cases with broadcasts not on innermost dim

cpu: x64: update Xbyak to 5.99

this closes oneapi-src#859

build: bumped version to v1.6.5

gpu: jit: disable mid-thread preemption

build: bumped version to v1.6.4

benchdnn: matmul: fix argument string