It's just a rendered still of a 3D scene used as background, the flames are sprites sitting on on top of it (notice they for not alter the lighting).
The player than moves on top of a simple lowpoly collision mesh that do not render and that mimics the traversable surfaces of the scene.
The scene camera is placed in the exact position that the image is rendered from relative to the collision mesh, as long as you keep the camera here and do not translate it you can
use a narrower FOV and rotate it to make it scroll with the player revealing only a subpart of the rendered background as you move around and explore the scene.
This last part is the only real trick to the setup as you are not using an isometric view for these.>>519427>cubemaps
This is like 10 years before cubemaps, also wth are you thinking using cubemaps for this?