Investigate usage of sparse vectors in GObservations::likelihood::eval method

The GObservations::likelihood::eval() method still shows up as consuming a non-negligible amount of computing time when analysis SPI data due to the usage of GMatrixSparse::operator+, GMatrixSparse::column() and GVector::non_zeros(). Similar to actions done under issue #9 (closed), using sparse vectors can speed thing up (see below).