[MLIR] Changes to fallback Batch implementation

- When copying the attributes currently we use the attributes for the source operation. Instead for better default operation coverage, we can do:
- If name contains "dimensions" and is a array attr, we offset each of the dimensions by `batchSizes.size()`
- If name contains "dimension" and is a integer attr, we offset that by `batchSizes.size()`
- When cloning an operation with no operands, we should instead clone the operation without updating anything and instead do a broadcast_in_dim on the result of the operation. We should clone the regions as we did before.