Introduce matmul-based IP for training

[densamoilov]

This PR introduces matmul-based IP implementation for training.

• When IP is created for forward_training plain layouts are always used to be consistent with the corresponding formats used for backward_data and backward_weights
• To support bias gradients for backward_weights case the matmul primitive has been extended with an internal feature that enables mamtul to reduce matrix A. The reduction is fused.
• Verbose was not extended to cover the new internal feature intentionally because having it in verbose doesn't provide any information that would be helpful, information about bias for IP is already enough.
• This matmul-based implementation for training is meant to be a full replacement for the existing brgemm-based one however, the latter one will stay available through the primitive descriptor iterator for some time but no further development is planned for it and it will eventually be removed.

TODO: post performance data.

[densamoilov]

make test
enable device_cpu
disable device_gpu