Mixing baked and real-time shadows and real-time performance

1) The whole idea behind dual lightmaps is to support this kind of workflow. Near the camera the dual lightmaps only store bounce lighting and the local lighting + shadows comes from the dynamic light. Far away, the dynamic shadows are disabled and the lightmaps switch to baked versions. Is there a reason dual lightmaps don't work for you?

2) The specifics vary depending on a lot of specifics, but usually the main cost is the draw calls for all of the shadow casting objects. You basically have to have one draw call per shadow casting object per light source that casts shadows from that object. Directional light shadows use cascaded shadows which basically act like a unique shadow casting light source per shadow slice. Basically, the former is probably more CPU intensive while the latter is maybe more GPU intensive, but on most platforms CPU time is harder to come by than GPU time.