Improve SIT to compose a batch tensor as a model input make up from a packet of different images

[sivanov-work]

Details:

• A new parameter override_model_batch_size was added several PRs ago. The purpose of this parameter was is to take a picture/binary and "stretch" it to fit a batched tensor as an input of a batched model providing that we could run and validate batched models
• The current PR enhance this idea and allows us to set distinguished images/binaries in a batched tensor input so that we could validate correctness of processing of different lines of batch during our inference invocations. SIT now uses "|" separator to discern different images inside input images array as an command line argument.

The batched images filing algorithm is the following:

1. If passed images amount is not enough to fill up an entire batch, a last image will be propagated and copied to remnant lines of N-batch.
2. If passed images array is larger than an batched tensor, then only N images will be used up and the rest of images will be thrown away

Example:

single-image-test --device CPU --network resnet50.xml --mode classification --top_k 5 --il NCHW --ol NC --iml NCHW --oml NC --ip FP16 --op FP16 --override_model_batch_size 3 --input "dog.bmp|00000001.JPEG|00000002.JPEG"

gives us

...
Load input #0 from [dog.bmp,00000001.JPEG,00000002.JPEG] as f16 [N,C,H,W] [3,3,224,224]
Fill up all input batch slices planes up to 3 with image data from the array: [0/3]
Fill input batch slices planes starting from index 1 up to 3 with image data from the array: [1/3]
Fill input batch slices planes starting from index 2 up to 3 with image data from the array: [2/3]

Tickets:

• E####145303

[ArtemySkrebkov]

I am not familiar with batching, but the changes look reasonable.

[DariaMityagina]

LGTM 👍