Separate linear operator from matrix trait

[milibopp]

For context, this idea is motivated by vbarielle/sprs#118.

The matrix trait is rather large at the moment:

pub trait Matrix: Sized + Clone + Mul<Self::Row, Output = Self::Column> {
    type Field: Field;
    type Row: FiniteDimVectorSpace<Field = Self::Field>;
    type Column: FiniteDimVectorSpace<Field = Self::Field>;
    type Transpose: Matrix<Field = Self::Field, Row = Self::Column, Column = Self::Row>;
    fn nrows(&self) -> usize;
    fn ncolumns(&self) -> usize;
    fn row(&self, i: usize) -> Self::Row;
    fn column(&self, i: usize) -> Self::Column;
    unsafe fn get_unchecked(&self, i: usize, j: usize) -> Self::Field;
    fn transpose(&self) -> Self::Transpose;

    fn get(&self, i: usize, j: usize) -> Self::Field { ... }
}

For many algorithms, access to the internal structure of the matrix is not necessary. It also does put some burden on trait implementors that may be possible to avoid.

A suggested solution would be to introduce a more abstract linear operator inspired by SciPy's interface with the same name as a supertrait.

pub trait LinearOperator: Sized + Clone + Mul<Self::Row, Output = Self::Column> {
    type Field: Field;
    type Row: FiniteDimVectorSpace<Field = Self::Field>;
    type Column: FiniteDimVectorSpace<Field = Self::Field>;
    type Transpose: LinearOperator<Field = Self::Field, Row = Self::Column, Column = Self::Row>;
    fn nrows(&self) -> usize;
    fn ncolumns(&self) -> usize;
    fn transpose(&self) -> Self::Transpose;
}

pub trait Matrix: LinearOperator {
    fn row(&self, i: usize) -> Self::Row;
    fn column(&self, i: usize) -> Self::Column;
    unsafe fn get_unchecked(&self, i: usize, j: usize) -> Self::Field;
    fn get(&self, i: usize, j: usize) -> Self::Field { ... }
}

It would be useful to include matrix multiplication there as well, but I have a feeling that this is hard to express in today's Rust. It also seems like a separate issue as the original matrix trait does not include this either.