Generally looks fine to me. But please find and fix the reason that it doesn't work for a Union operation.

Hmm, in my testing union is working fine.

For the record I'm going to post some performance information I talked about with Howard in chat:

My discovery was the this patch is actually slower when there are multiple input meshes.
This is because boolean has an optimization to avoid calculating self intersections when there are only two input shapes (using two BVH trees rather than a single larger one).
Since realizing the instances is much faster than boolean intersections, realizing instances for each input geometry component avoids that situation and is much faster.

This node is still about 40% faster when there are only two input meshes, and it's the same as the implementation for the modifier, so Howard and I decided it would be best to move forward anyway.
The boolean code can be optimized for more input meshes in the future, and there's already a workaround for situations where it matters anyway-- just realize instances first with a different node.

WITH PATCH APPLIED
Timer 'geo_node_boolean_exec' took 162.446 ms # This just gathers instanced meshes and their transforms from the geometry set
  Timer 'direct_mesh_boolean' took 162.427 ms
    Timer 'meshes_to_imesh' took 1.73373 ms
    Timer 'boolean_mesh' took 157.892 ms
    Timer 'imesh_to_mesh' took 0.800372 ms

WITHOUT PATCH APPLIED
Timer 'geo_node_boolean_exec' took 54.879 ms
  Timer 'geometry_set_realize_instances' took 230 ns # This geometry set's mesh is not an instance, so this doesn't do anything
  Timer 'geometry_set_realize_instances' took 0.233303 ms # This combines the 5 input meshes from the collection into one larger one
  Timer 'mesh_boolean_calc' took 54.591 ms # Then convert to BMesh
    Timer 'boolean_mesh' took 46.4092 ms # And finally call boolean

MODIFIER (for reference)
time end   (__func__): 0.175889  source/blender/modifiers/intern/MOD_boolean.c:716
  Timer 'direct_mesh_boolean' took 175.86 ms # This is doing the same thing as the "with patch applied" results
    Timer 'meshes_to_imesh' took 2.56968 ms 
    Timer 'boolean_mesh' took 170.486 ms
    Timer 'imesh_to_mesh' took 0.784861 ms

Sorry about the intrusion but it would be awesome if we got to see the fast solver with geometry nodes, the exact solver is not very useful in many situations with complex geometry (at least when realtime feedback is important) - it slows down the viewport too much.

Some other long existing feature request among hard surface artists (related to boolean operations) that I would also like to lift up is that we are missing a way of assigning attributes to the cut out geometry (as well as the boundary loop) - so that one could perform operations like bevel or inset (if we are getting inset that is) on cut out geometry.