Add sparsity optimizations for matrix folding

[Fantoni0]

The mova_matrix.rs file in the feat/matrix_folding branch implements the basic folding protocol for MOVA-style matrix folding. But if we know that the first instance ($$I_1$$) is going to be sparse and using ternarized weights (-1, 0,1), we can apply further optimizations techniques. We should:

• Either tag the current implementation as "matrix-folding-simple" or start a new optimized implemenation in a new branch.
• Review how the folding is done: https://www.notion.so/nethermind/Checking-the-cross-term-of-the-protocol-5-when-we-use-sparsity-of-the-first-instances-183360fc38d080e89a8df162ff561c33
• Bring the SparseMatrix implementation and Matrix wrapper from the FLI implementation, and change the witness types.
• Study if the sparsity can also be leveraged when computing MLEs.
• Add a small benchmark to compare basic folding and optimized folding.