The most preferred option 3, since in this case it is assumed that each element will have a class that meets almost all of the major methodologies, and dearly loved me SMACSS.
The performance of the cascade is affected directly, because, firstly, increasing the size of the css file blocking the rendering, and second, the reading of the selectors and the search in the tree right to left, i.e. in the embodiment 1 will first find all references and images, and then collected heap violents ones that are in the hat. But ideally, of course, will be something like this:
No cascade (the size of the output css is minimized), the loss of productivity in the sample is also minimized.