x64: updates for brgemm kernel and avx2 brgemm convolutions

[ankalinin]

This request contains a few general brgemm code changes and some performance updates for avx2:

• brgemm: introduce brgemm_desc_finalize function : this function must be called after all the parameters of the brgemm descriptor are set. As result brgemm blocking is called only once having all needed information
• brgemm conv: update to reduce the number of generated kernels: small update for brgemm non-amx convolutions
• brgemm: update non-amx blocking: more careful distribution of available vmm registers. It's especially important for avx2. It also determine which order of data loading in microkernel to choose - n_bcast_1_load or 1_bcast_n_load .
• brgemm kernel: update gemm microkernel: just update the kernel code
• brgemm kernel: reduce number of vpad variants: it reduces the size of generated kernel
• brgemm conv: knob for brgemm microkernel decomposition: we may want to tune microkernel decomposition
• brgemm conv: update loop nest - move loop by ic_chunks to the top: we usually don't use blocking by input channels. If we do it makes more sense to have a corresponding block loop at the top level
• brgemm conv: introduce loop order gcndhw: for some shapes such loop order makes sense in terms of performance
• brgemm conv: update cache constants and ur estimation: update platform values used in convolution blocking selection
• brgemm conv: heuristic for big convolution on avx2

Performance testing by openvino on MTL. Here are a ratio brgemm to jit implementation for rls-v3.5 and for this brgemm update over v3.5:

[ankalinin]

make test
disable test_device_gpu
disable build_gpu_runtime_ocl
disable build_gpu_runtime_sycl

[Sqvid]

@Radu2k could you have a look at the aarch64 changes? Thanks

[Radu2k]

LGTM, also no regressions in terms of performance. Thanks!