inconsistency in _solve_trig when it comes to CRootOf

[gschintgen]

In [5]: solveset(tan(x/6) - cos(x/2))
Out[5]:
⎧                ⎛       ⎛ 8      7       6      5       4      3       2             ⎞⎞        ⎫   ⎧                ⎛
⎨12⋅n⋅π + 12⋅atan⎝CRootOf⎝x  - 2⋅x  - 16⋅x  - 6⋅x  + 30⋅x  - 6⋅x  - 16⋅x  - 2⋅x + 1, 0⎠⎠ | n ∊ ℤ⎬ ∪ ⎨12⋅n⋅π + 12⋅atan⎝CRoot
⎩                                                                                               ⎭   ⎩

  ⎛ 8      7       6      5       4      3       2             ⎞⎞        ⎫
Of⎝x  - 2⋅x  - 16⋅x  - 6⋅x  + 30⋅x  - 6⋅x  - 16⋅x  - 2⋅x + 1, 1⎠⎠ | n ∊ ℤ⎬
                                                                         ⎭

_solve_trig1 catches RootOf in results and returns NotImplementedError. Such handling should also be added to _solve_trig2.

[gschintgen]

Wait, I'll take that back. It's the other way round: maybe _solve_trig1 shouldn't prematurely abort if RootOf objects appear. After all, those solutions should be correct and they could still be useful. For example they could be evalf()ed by the user.
(But that requires ImageSet.evalf to work properly. See #17382)