C++: Reduce memory pressure from getInstruction

[MathiasVP]

Since we merged #12900 the Foundations teams internal performance measurements have been OOM'ing when analyzing Wireshark because they're running their experiments with ~6GB of RAM (unlike our DCA analysis of Wireshark which uses a much larger runner).

The analysis was OOM'ing because of the use of the shortestDistances HOP. This use was introduced in 8368c37 to get around an issue where Eclipse was OOM'ing because the log file was too large 😅.

This PR reverts 8368c37 and replaces the transitive closure of adjacentInBlock (which also was OOM'ing when I tried it locally) with an EquivalenceRelation-based approach. This succeeds even with 5GB of memory 🎉.

We do, however, now have 148647 iterations of getMemberIndex 😱. So we need to run DCA (and maybe even MRVA?) to check if this is a change we want to make.

[MathiasVP]

Hmmm... Apparently, this breaks some of our tests 🤔. I'll investigate.

Edit: Fixed in 57cc316.

[rdmarsh2]

Stage timings on DCA look pretty noisy. I'm OK with the code but we should look into that.

[MathiasVP]

Stage timings on DCA look pretty noisy. I'm OK with the code but we should look into that.

Sure, we can do that. Note that the stage timings behavior is actually quite consistent with our observations. In all of the projects, the CppType::CppType#1#IRConstruction::Raw#14#Operand#3#SSAConstruction::Cached#7#TOperand::Internal#3 stage is the one with the largest regression (which is most likely the stage that evaluates getInstruction).

[MathiasVP]

Yeah, the query-specific stage timing alerts flagged up here is just noise.