perf(state-res): use roaring bitmaps for faster auth difference computing

[gamesguru]

One of the few magical times where adding more lines of code makes things run significantly faster.

Rapidly computes set unions and intersections with hardware-accelerated bitwise operations, keeping the L1/L2 cache hot throughout. Sometimes this means performing comparisons or operations on up to 256 graph nodes per clock cycle.

---

Future work / Next steps

Ultimately, this would further greatly benefit from a refactor which enabled loading the full auth chain into memory at once (if the machine's throughput would allow).

That's when you would start to see large evals happen in a tiny fraction of a second, because you could compute full topological sweeps and pre-compute transitive auth closures virtually at the silicon limit.

---

EDIT: About to push some commits that read directly from DB into roaring data types, should achieve similar speedup as this benchmark:

Starting Comparative Performance Benchmark...

--- Benchmark with size: 100 elements across 20 sets ---
Origin/Main (BTreeMap Counts):      2.856805795s (avg: 142.84µs)
Cached/Pre-computed Roaring:      189.769309ms (avg: 9.488µs)
Speedup of Cached Roaring over origin/main: 15.05x

--- Benchmark with size: 2000 elements across 20 sets ---
Origin/Main (BTreeMap Counts):     89.105965617s (avg: 4.455298ms)
Cached/Pre-computed Roaring:        4.187627498s (avg: 209.381µs)
Speedup of Cached Roaring over origin/main: 21.28x

[gamesguru]

@coderabbitai review

[gamesguru]

@cubic-dev-ai review