JS runtime: Use classes for Ref and Arena instead of closures

[jiribenes]

Previously, we:

1. for each var, we allocated a new closure for its own set method
2. for each RESET, we allocated two closures for their fresh and arena methods

... what if we, uh, didn't do that?

---

This could have been resolved both with prototypes and classes, but I chose to use classes since intent is important and classes are better recognisable by tooling. (and by JSDoc which is relevant for #1218)

The benchmark results are pretty strong: geomean is like a 4-5% speedup on my machine, and all benchmarks got faster, yet!

These are the biggest changes (ran with hyperfine, warmup = 5, N ≥ 10, inputs from the config_js.txt file), everything else is negligible as per the Welch t-test (using mean, stddev & number of runs; the tests where hyperfine complained about outliers are deemed as not significant):

Benchmark	main after #1330	this PR	Δ
mandelbrot	177 ms	108 ms	−38.8%
sieve	150 ms	134 ms	−10.6%
bounce	120 ms	114 ms	−5.2%
storage	690 ms	672 ms	−2.5%
product_early	248 ms	241 ms	−2.7%
triples	375 ms	366 ms	−2.4%
dyck_one	404 ms	385 ms	−4.8%
number_matrix	357 ms	344 ms	−3.7%
financial_format	294 ms	284 ms	−3.4%

[jiribenes]

Explicitly setting stack: null seems to help V8 so that it knows it's always the same "type".
Same for the one created in RESET.

[jiribenes]

I think this is largely uncontroversial, so I'll merge this -- we can always revert later before the release if the Effekt Plots decide to go into an undesirable direction.