Implement MaterializeEncoding for conv ops on GPU with IGEMM

Once #17605 is implemented, we can start making use for these encodings to materialize for certain conv implementation strategies. In particular, on GPU targets where we want to use the IGEMM strategy, this benefits from the C dimension being innermost, so we can materialize the encoding there are a transposition (or a special case of pack op) to achieve that.

Start by reviewing the prerequisite work in issue #17605 and then trace how convolution encodings are represented for GPU targets using the IGEMM strategy. Done means MaterializeEncoding supports these convolution operations and materializes the encoding so the C dimension is innermost, using a transpose or pack operation as appropriate.

Written by the indexing model from the issue text.