src: cpu: aarch64: add support for s8:s8:s8 in ACL lowp matmul

[michalowski-arm]

Performance results:

matrix scale | speed-up

 128x128   |  x0.82
 256x256   |  x2.63
 512x512   |  x33.3
1024x1024  |  x60.7
2048x2048  |  x60.3

To select the correct s8->s8 ACL kernel, we need to send
all quantization info at configuration but oneDNN does
not make these available until execution. This change
goes around this issue by first performing s8->f32 matmul
and then requantizing back to s8.

[jondea]

This looks great, thank you @michalowski-arm. This has also been reviewed internally.

[jondea]

If there are no more comments, would it be possible to get this merged please? The failures look common

[vpirogov]

@jondea, the failures are caused by an MSVC bug.

Thanks for the code review!