[FlashMatrix]: allow local matrices to be resized when evaluating a DAG

[zheng-da]

Currently, we only allow local matrices in a DAG to be resized only once. Maybe it's necessary to have every virtual local matrix to make such a decision.

If this is the case, every virtual local matrix needs to count the number of elements being used, instead of the number of times the local matrix being used, before dropping the local matrix.