Try to improve the performance issue described at #587.
The issue is that Mathlib contains thousands of theorems where the
associated key for the discrimination tree is just
`Key.other`. The indexing is not effective for them. This happens because
1- Lambda expressions are indexed using `Key.other`. The
discrimination tree mainly focus on the first-order structure.
2- It unfolds reducible constants when inserting and retrieving
entries. The motivation is that users expect simp theorems to fire
modulo reducible constants.
Then, we have many theorems such as
```lean
map ?g ∘ map ?f = map (?g ∘ ?f)
```
when we expand the function composition on the left-hand side, we get
```lean
fun (x : List ?α) => map ?g (map ?f x)
```
Which is indexed as `Key.other`.
We should not avoid the `Array`s in the discrimination tree nodes
If the index is working effectively, these arrays are all very small.
In this commit, we try to address the problem by using a different
approach. When processing the root of a pattern, we interrupt
reduction as soon as the we hit something that would be indexed
as `Key.other`. Note that, in Lean 3, the root of a pattern also
receives special treatment.
