And as a pleasant side-effect, somewhat improved performance when referencing
local vars as well :)

- Make integer equality check of `Number`s cheaper by not delegating
  to Compare unless required
- Fix mistake where IndexResult structs never got put back into the
  `sync.Pool` they came from. Oops!
- Slightly cheaper `numbers.range` by avoiding allocating an error
- Add `*eval.unknownRef` method to avoid allocations from `any` boxing
  of refs

I've also added a benchmark to easily try compare the costs of these two lookup types,
and I'm pretty happy with the results:

**main**
```
global_ref-12         	   20461	     57639 ns/op	   24473 B/op	     903 allocs/op
local_var-12          	   42226	     28495 ns/op	   13363 B/op	     410 allocs/op
```

**change**
```
global_ref-12         	   25377	     46168 ns/op	   14627 B/op	     496 allocs/op
local_var-12          	   46336	     25671 ns/op	   11488 B/op	     300 allocs/op
```

Fixes open-policy-agent#7654

While there are still potential improvements to be made, and referencing a global rule
is still more expensive than a local var, the difference is now small enough that the
global -> local doesn't feel warranted for anything but the most demanding requirements.
For that reason, I consider this fixed. The benchmarks are great for anyone looking to
improve this further :)

Signed-off-by: Anders Eknert <anders@styra.com>