graph: backend: dnnl: support permute for scale and zps inputs

[wzt1997]

Description

We recently supported the compressed SDPA patterns which incorporate scales and zero points as inputs. Also, we aim to support the problem with a Key shape of ( N, H, S, D ) with transpose_b=true for the QK MatMul. For such case with transposed MatMul, the graph library adds permute ops prior to the pattern input, including the scale and zp tensors. To be more specific, for the following woq MatMul with tranpose_b=true:

After the graph compilation, we will have the graph as follows. The permute ops will not touch the physical memories but only change the mds.

However, we faced a issue that DNNL backend always regards scale and zp tensors as abx format, leading to failing results, hence we need to set the tag explicitly as abx and execute extra reorders. After the change, the graph after compilation is like:

[wzt1997]

make test
enable benchdnn_nightly
disable benchdnn_all
enable benchdnn_graph

[wzt1997]

make test
enable benchdnn_nightly
disable benchdnn_all
enable benchdnn_graph

[wzt1997]

make test
enable benchdnn_nightly
disable benchdnn_all
enable benchdnn_graph

[wzt1997]

During the validation of dynamic per-channel quantization cases, especially for those with zps and transposed matmul, some issues were discovered. So I pushed 2 new commits for enhancing the handling for dynamic per-channel quantization cases in the PR, please review again:

1. Apart from per-group quantization and single reorder primitive, the backend always accepts mask=0 for zero points, hence improved the op def check for dynamic quantization ops to allow zp tensor of which nelems=1 for per-channel quantization.
2. Added corresponding compressed SDPA case with per-channel quantization, and the transposed test case.

[wzt1997]

make test
enable benchdnn_nightly
disable benchdnn_all
enable benchdnn_graph