[CPU][ARM] ACL int8 Convolution executor shouldn't support non-i32 bias

[alvoron]

Details:

• Before the fix ACL int8 convolution executor was chosen for fp32 bias case. Previous type mapping forces fp32 to int32 conversion, which led to accuracy degradation
• Type mapping has been fixed to accept i32 bias only
• If bias is not i32 then such case is handled by dnnl executor. To do that the order of ARM executors has been changed: int8 executor first, default dnnl executor next.

Tickets:

• CVS-175513

[EgorDuplensky]

Do we know why the bias is i32 in the first place?
I think this is because from ACL point of view we are supposed to quantize bias (apply activation and weights scale on it).
The question is when to do it.
Ideally it should be done by ov pass (replace 'conv_int8+bias_f32' with 'conv_int8+bias_i32' by applying scales on bias const)
Thus, we will not need to fallback to slower oneDNN implementation.