[CPU] Fill the ellipsis mask in StridedSlice node to match the input rank

[maxnick]

Details:

Fix the ellipsis mask preprocessing in the StridedSlice node to avoid out of bounds memory access.

Tickets:

• 128638

[maxnick]

@a-sidorova , could you please review?

[a-sidorova]

In general LGTM.

But seems like...

• as we discussed offline, the current specification doesn't cover ranks of masks, paddings- and other cases. I think we have to prioritize the ticket 90128
• the current implementation is based on assumption that the masks are mapped on input shape (or output??) starting from first element. If masks and shapes rank are different, we expended masks by empty values. But it's not described in the spec as well. Our implementation is written using different assumptions but not the spec. Moreover the reference implementation is different from plugins (the part of masks and paddings, the work with normalization). It might leads to new bugs. I'm not sure that aligning with the reference is good idea as main solution. Because reference may contain bugs as well until we don't have the fully described specification (I remember that I tried to align our impl with the reference code but I see that some parts in the core have been updated). So I think we have to update the specification to avoid many misunderstandings and possible bugs.

[maxnick]

In general LGTM.

But seems like...

• as we discussed offline, the current specification doesn't cover ranks of masks, paddings- and other cases. I think we have to prioritize the ticket 90128
• the current implementation is based on assumption that the masks are mapped on input shape (or output??) starting from first element. If masks and shapes rank are different, we expended masks by empty values. But it's not described in the spec as well. Our implementation is written using different assumptions but not the spec. Moreover the reference implementation is different from plugins (the part of masks and paddings, the work with normalization). It might leads to new bugs. I'm not sure that aligning with the reference is good idea as main solution. Because reference may contain bugs as well until we don't have the fully described specification (I remember that I tried to align our impl with the reference code but I see that some parts in the core have been updated). So I think we have to update the specification to avoid many misunderstandings and possible bugs.

СС @mitruska