[CPU] Add test case of setting model_distribution_policy to TENSOR_PARALLEL

[sunxiaoxia2022]

Details:

• add test case of model_distribution_policy is TENSOR_PARALLEL
• Enable "ov::hint::ModelDistributionPolicy::TENSOR_PARALLEL" on single socket will create duplicated sub stream for functionality validation

Tickets:

• CVS-162977

[dmitry-gorokhov]

It has to be usual functional subgraph test, which already have all the infra for output comparision and inputs filling.
Besides that we have to cover cases like matmul with compressed weights.
It also worth to verify execution graph to make sure TP was actually applied

[sunxiaoxia2022]

Ok, I added test case about TP to mat_mul.cpp and shared_matmul_weights_decompression.cpp. Please have a look.

[EgorDuplensky]

Details:

• add test case of model_distribution_policy is TENSOR_PARALLEL
• Enable "ov::hint::ModelDistributionPolicy::TENSOR_PARALLEL" on single socket will create duplicated sub stream for functionality validation

Tickets:

• CVS-162977

@sunxiaoxia2022 Does it mean we allow setting TP on any single socket machine even for Release package? Are there any potential performance degradations because of that? I understand that is done just to enable testing of the feature but having it always available in API can have potential negative impact.

[wangleis]

Details:

• add test case of model_distribution_policy is TENSOR_PARALLEL
• Enable "ov::hint::ModelDistributionPolicy::TENSOR_PARALLEL" on single socket will create duplicated sub stream for functionality validation

Tickets:

• CVS-162977

@sunxiaoxia2022 Does it mean we allow setting TP on any single socket machine even for Release package? Are there any potential performance degradations because of that? I understand that is done just to enable testing of the feature but having it always available in API can have potential negative impact.

@EgorDuplensky Yes, it allow setting TP on single socket machine even for release package. Since TP is disabled by default, this change will not impact default behavior.

If the user mistakenly sets TP for real inference on a single-slot platform, negative impact is expected.