Reading sparse matrix of unknown type

[ExpHP]

The sparse module was designed a bit hastily and it is only after trying to use it in another project that I noticed the following discrepancy:

NpyFile allows an NPY file to be attempted to be parsed using multiple dtypes, yet the sparse types demand knowing the type upfront.

Adding more functionality to the sparse types seems unwise. (I'd hate for those docs to become even more inscrutible). I wish there was a nice in-between solution, where the user could use NpzWriter directly, but had access to functions that somehow made it easier to correctly read a sparse matrix (automatically reading the correct filenames with the correct dtypes for indices). I'm not sure what that would look like.

[ExpHP]

Note: I briefly considered removing the data field from sparse types and letting the read function return an NpyFile for the Data. (this would also solve the incorrect derived Clone bound on these types)

// was:
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct CooBase<T, Data, Indices>
where
    Data: Deref<Target=[T]>,
    Indices: AsRef<[u64]>,
{
    pub shape: [u64; 2],
    pub data: Data,
    pub row: Indices,
    pub col: Indices,
}

impl<T: Deserialize> Coo<T> {
    pub fn from_npz<R>(npz: &mut NpzArchive<R>) -> io::Result<Self>
    where R: io::Read + io::Seek { ... }
}

// now:
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct CooBase<Indices>
where
    Indices: AsRef<[u64]>,
{
    pub shape: [u64; 2],
    pub row: Indices,
    pub col: Indices,
}

impl Coo {
    pub fn from_npz<R>(npz: &mut NpzArchive<R>) -> io::Result<(Self, NpyFile<_>)>
    where R: io::Read + io::Seek { ... }
}

But I'm not sure what the write functions would look like. There is a particularly weird thing with Dia because the second dimension of its data cannot be determined without knowing all of the data. I guess a length field is needed.