Optimize Thread.detect_recursion

[eregon]

We should optimize Thread.detect_recursion, e.g., by moving some of it to Java, using EconomicMap instead of {}.compare_by_identity which forces computing object_id.

That should also do the optimization of not computing the identity hash until needed if few objects are tracked.

We should also avoid the block if possible as that creates more indirections.

From #1992 (comment)

[chrisseaton]

@LillianZ has some benchmarks and works in progress towards solving this.

[chrisseaton]

We've got some significant work on this now - will merge as much of it as we can soon but please don't refactor anything to do with detect_recursion until then thanks.

[norswap]

Re: #2189 (comment)
I guess that idea is dead in the water, but just to be very sure: there is no way to leverage some kind of thread-local storage in a way that would be faster than optimizing our own maps, right?

[eregon]

Follow-ups for this after #2189:

• Avoiding an extra lookup in DeleteLastNode: Detect recursion #2189 (comment) (done in Optimize Thread.detect_recursion #2011 too)
• Make Truffle::ThreadOperations.detect_recursion a primitive so it's always split, in a cheaper way than using Truffle splitting of CallTargets: Detect recursion #2189 (comment)

[eregon]

Related optimization in CRuby, to avoid using the recursion logic when we know we'd call Kernel#hash: ruby/ruby#3987