That looks wrong to me. Why does the ellipse have diagonal lines and the circle/square doesn't? It seems to me like nothing should have diagonal lines in this geometry.
Consider creating taxicab shapes on graph paper. If you color in cells that satisfy the definition of a circle or ellipse, you're right that you won't see diagonal lines, only colored squares. If, however, you shrink the size of the grid further and further until it's infinitesimally small, those points will take on the appearance of diagonal lines.
This is an interesting idea, but I'm not sure how it helps understanding. As you point out in your first paragraph, taxicab geometry differs from ordinary geometry only in the way that it measures distances (I shy away from saying 'lengths', particularly of curves, because it's not clear to me that the Euclidean theory of rectifiable curves has a nice analogue in taxicab geometry); and distance is a point-point property, not a property of cells. How could one decide whether or not a cell satisfies the definition, except by picking a point in it? (That's not a rhetorical question.)
(Of course, as you say, the error involved can be made 'small', in some sense, by making the grid suitably fine; but at that point, with such a simple metric as the taxicab one, it's not clear to me what you're gaining over just looking at the entire plane all at once.)
Edit: For clarification, the "points" in both the Euclidean and taxicab case can be represented in Cartesian coordinates (e.g. (x, y)). What is different is how you define the distance between those points. In Euclidean geometry, the distance is r = Sqrt((x - x')^2 + (y - y')^2), whereas in taxicab geometry the distance is r = |x - x'| + |y - y'|.
Indeed, if one were trying to understand, say, the locus of e^(x + y) = x^2 + y^2, which is an unfamiliar shape then I would say by all means to discretise it; that's what visualisation software would do, after all.
However, conic-section analogues defined via linear constraints on distances will, in the taxicab metric, always consist of unions of line segments, and it seems to me that discretisation is likely to hurt, not help, visualisability of such shapes.