graph: backend: dnnl: make matmul use any layout format only for constant cases

[Jiexin-Zheng]

Description

Create matmul src and weight mds with any layout only when they are constant

Fix MFDNN-13714

Performance improvements

benchdnn graph

HW	mha cases speedup
PVC	1.21x
BMG	1.17x
Arc-A770	1.30x
SPR	1.01x

Pytorch GPU model

model	dtype	HW	inductor speedup	eager speedup
GPT2ForSequenceClassification	fp32	PVC	1.06203	1.05513
GPT2ForSequenceClassification	fp16	PVC	1.39628	1.35299
GPT2ForSequenceClassification	bf16	PVC	1.40592	1.35395
PLBartForConditionalGeneration	fp32	PVC	1.07212	0.99772
PLBartForConditionalGeneration	fp16	PVC	1.03491	1.00943
PLBartForConditionalGeneration	bf16	PVC	1.02930	1.00978
MegatronBertForCausalLM	fp32	PVC	1.01970	0.99272
MegatronBertForCausalLM	fp16	PVC	1.03874	0.99610
MegatronBertForCausalLM	bf16	PVC	0.99199	1.00457
hf_GPT2	fp32	PVC	1.03786	1.03452
hf_GPT2	fp16	PVC	1.44831	1.40425
hf_GPT2	bf16	PVC	1.44748	1.40106
llama_v2_7b_16h	fp32	PVC	1.00268	1.00336
llama_v2_7b_16h	fp16	PVC	1.01445	1.00231
llama_v2_7b_16h	bf16	PVC	1.06454	1.01447

[Jiexin-Zheng]

make test
set test_scope=NIGHTLY
disable benchdnn_all
enable benchdnn_graph

[Jiexin-Zheng]

make test
set test_scope=NIGHTLY
disable benchdnn_all
enable benchdnn_graph

[Jiexin-Zheng]

make test
set test_scope=NIGHTLY
disable benchdnn_all
enable benchdnn_graph

[Jiexin-Zheng]

make test
set test_scope=NIGHTLY
disable benchdnn_all
enable benchdnn_graph