Previously:
If the `rfl` macro was going to fail, it would:
1. expand to `eq_refl`, which is implemented by
`Lean.Elab.Tactic.evalRefl`, and call `Lean.MVarId.refl` which would:
* either try kernel defeq (if in `.default` or `.all` transparency mode)
* otherwise try `IsDefEq`
* then fail.
2. Next expand to the `apply_rfl` tactic, which is implemented by
`Lean.Elab.Tactic.Rfl.evalApplyRfl`, and call `Lean.MVarId.applyRefl`
which would look for lemmas labelled `@[refl]`, and unfortunately in
Mathlib find `Eq.refl`, so try applying that (resulting in another
`IsDefEq`)
3. Because of an accidental duplication, if `Lean.Elab.Tactic.Rfl` was
imported, it would *again* expand to `apply_rfl`.
Now:
1. Same behaviour in `eq_refl`.
2. The `@[refl]` attribute will reject `Eq.refl`, and `MVarId.applyRefl`
will fail when applied to equality goals.
3. The duplication has been removed.
