Make nodes() return the correct indices when passed a view of a Field

[tomchor]

Resolves #4729

In addition to improving the behavior of nodes(), this PR, together with #4730, makes it possible to write different slices of a Field (or several Fields) into one NetCDF file.

[glwagner]

Should nodes return a KernelFunctionOperation? If it did, would this PR be easier?

[tomchor]

Should nodes return a KernelFunctionOperation?

Not sure. I imagine this being more cumbersome for the end user, but possibly reducing code and makings things more flexible.

If it did, would this PR be easier?

A little, yes. But I don't think enough as to motivate this change by itself.

[glwagner]

After #4805, it will be pretty trivial to return KernelFunctionOperation for nodes. Do you want me to wait for this PR to do that, or should I do that before this PR so you don't have to worry about CPU vs GPU for nodes?

[glwagner]

Basically if nodes return KernelFunctionOperation, then the way to get correct indices will be the same as building a Field and passing indices to the Field constructor. We won't need any "special" functionality for windowing the nodes, compared to windowing Field.

[tomchor]

After #4805, it will be pretty trivial to return KernelFunctionOperation for nodes. Do you want me to wait for this PR to do that, or should I do that before this PR so you don't have to worry about CPU vs GPU for nodes?

You can go ahead and make nodes return KernelFunctionOperation first!