Generalize ensemble Kalman algorithms to work with complex learnable parameters

For users that want to learn complex parameters (e.g., coefficients within an FNO), it would be useful to generalize the algorithms to directly allow them to work with complex numbers.

This would require modifying the covariances to use complex conjugates, and generalizing input types since

```
julia> b = sqrt(Complex(-2))
0.0 + 1.4142135623730951im

julia> isa(b, AbstractFloat)
false
```

Start by locating the ensemble Kalman algorithm implementations and any covariance or input-type tests. Determine the scope of complex-number support, including conjugate-based covariance calculations and accepted parameter types; done means the algorithms handle complex learnable parameters without breaking existing real-valued use.

Written by the indexing model from the issue text.