feat: review padding and type conversions in rollup

[MirandaWood]

Now that we have blobs, we also have some root public inputs in the form of BigNums. The numbers are too big to fit into a field, and we don't yet hash all public inputs, so for now we convert bytes32 to and from these numbers:

  function _bytes32ToBigNum(bytes32 _input)
    internal
    pure
    returns (bytes32 firstLimb, bytes32 secondLimb, bytes32 thirdLimb)
  {
    firstLimb = bytes32(uint256(uint120(bytes15(_input << 136))));
    secondLimb = bytes32(uint256(uint120(bytes15(_input << 16))));
    thirdLimb = bytes32(uint256(uint16(bytes2(_input))));
  }

Another public input is a 48-byte compressed point, which (to fit into 2 fields) is encoded as a 31 byte and 17 byte pair:

          // To fit into 2 fields, the commitment is split into 31 and 17 byte numbers
          // TODO: The below left pads, possibly inefficiently
          // c[0]
          publicInputs[offset++] = bytes32(
            uint256(uint248(bytes31(_submitArgs.blobPublicInputs[blobOffset:blobOffset += 31])))
          );
          // c[1]
          publicInputs[offset++] = bytes32(
            uint256(uint136(bytes17(_submitArgs.blobPublicInputs[blobOffset:blobOffset += 17])))
          );

The padding/conversions may be ineffecient and should be reviewed.