Optimize ripper bounds, and one other small perf tweak

[Earlopain]

Basically a port of ruby/ruby@c45f781 into ruby

It's quite effective at ~97% hit rate for me. Speeds it up from ~6.77x slower to only 4.07x slower.

For the lexer on_sp it also gives a bit of an improvement: 1.04x slower to 1.10x faster

I guess the class may be universally useful but for now I just made it nodoc. I also didn't look much into the window, it seemed good enough.

---

Other commit is a small hack to skip redudant pack/unpacking of locations. Prism by default packs locations into a 64bit integer and later extracts start offset / length to create the location. But we know we access nearly all of the locations anyways, so we can request a frozen result to compute them eagerly. It gives about a 4% speed boost.

---

Benchmark from 2ea8139

[kddnewton]

I don't think I want that generally in parse result. Can you put that into the ripper translator? (Just not sure if we truly want to support that for everyone going forward.)

I see you must have determined unshift wasn't as fast as reassigning and splatting the array. That really surprises me, are you sure that's the case?

[Earlopain]

I don't think I want that generally in parse result. Can you put that into the ripper translator?

Yeah, I can move it. I had it nodoc but that's fine too.

I see you must have determined unshift wasn't as fast as reassigning and splatting the array. That really surprises me, are you sure that's the case?

It's probably not. But it also only happens for what ripper considers void statements. The change is only to make 3e8c264 work, and overall it comes out positive.