Is there a way to see (on each graphics card), how many primitives that are optimal to send in each call?

The D3DCAPS9 structure only contains the maximum number of primitives each card can draw. Generally, what game developers will do is run their app on each card and determine the engine settings that produce the best performance. This is all entered into a database. Then, at installation, the app looks up your card and selects the appropriate out-of-the-box settings.

That whole approach is unlikey to work for you, since you need to test on each of the cards. However, you can always provide some option that the users can change.

AFAIK the batch size is more affected by D3D overhead than by the card. You should just make it large (thousands of vertices) if possible. Most of the card specific optimisation is done at shader level or by removing heavy and advanced effects.