[WIP][Pruning algo] Filter counting fix for linear layers

[daniil-lyakhov]

Changes

Redundant input shape param was removed from count_flops_and_weights_per_node function. count_filters_num now always takes into account pruned linear layers. count_flops_and_weights_per_node now correctly computes statistics for model with pruned linear layers.

THIS WILL CHANGE PRUNING LEVEL OF MODELS WHICH CONTAINS PRUNABLE LINEAR LAYERS, PRUNED BY FLOPS

Reason for changes

collect_input_shapes function isn't protected from case when input edges don't exist (unlike collect_output_shapes, which process nodes without output edges correctly). Turns out that output of collect_input_shaped used only in count_flops_and_weights_per_node function and could be easily replaced by linear layers in_features and out_features attributes. Besides count_flops_and_weights_per_node function didn't take into account pruned rows/cols of linear layer, which was fixed by getting parameters from input_channels and output_channels params.

Related tickets

86889

Tests

TODO