[GPU] Baseline InferRequest changes for dynamic shapes support

[andrew-k-park]

Details:

• Divide InferRequest into InferRequest with API 2.0 and InferRequest with legacy
• Remove cldnn::layout depedency from create_device_blob
• Add reinterpret_device_blob and ov::PartialShape constructor in layout temporarily

[yeonbok]

Lets check this with networks benchmark

[andrew-k-park]

/azp run openvino-win

[azure-pipelines[bot]]

Azure Pipelines successfully started running 1 pipeline(s).

[andrew-k-park]

I still see some legacy stuff in new request impl and looks like handling of dynamic shapes is missing as well. Example from GetBlob method:

        if (isDynamic) {
            // code below is related to legacy dyn batch impl. Not handling for new dyn shapes flow.
            if (m_graph->GetMaxDynamicBatchSize() > 1) {
                SizeVector outDims = data->getTensorDesc().getDims();
                outDims[m_graph->GetOutputDynBatchDims()[name]] = m_curBatch;
                data->getTensorDesc().setDims(outDims);
            }
        } else {
            checkOutputBlob(data, name, foundOutput);
        }

Please review each method carefully and ensure that legacy features handling is removed from new impl and dynamic shapes processing is added.

Cleaned up the remaining legacy feature handling. Please review again to see if there is any dynamic shape processing that needs to be added.

[vladimir-paramuzov]

@andrew-k-park Could you try to add some simple functional test for dynamic case and check infer request behavior in it? Maybe activation primitive will work...