Experiment: Lazy arena allocation on JS

[jiribenes]

On LLVM, we allocate the arena for mutable variables only on demand. Why not do the same in JS?

---

The implementation is a bit ugly, I think this could be improved if we merge #1332 first.
I left in a lot of comments so that I understand what the heck is going on, these should be tidied up a bit before merging.

Speedup in total is some 3-4% (geomean), everything that sometimes allocates got much faster.
Below are the significant 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	177.1 ms	129.3 ms	−27.0%
parsing_dollars	1234.4 ms	1129.0 ms	−8.5%
triples	374.9 ms	357.2 ms	−4.7%
iterator	115.4 ms	112.3 ms	−2.7%
sieve	150.2 ms	146.6 ms	−2.4%
word_count_ascii	576.8 ms	563.1 ms	−2.4%
erase_unused	242.0 ms	237.3 ms	−1.9%
word_count_utf8	371.4 ms	364.3 ms	−1.9%
nqueens	199.0 ms	196.8 ms	−1.1%
factorial_accumulator	2354.2 ms	2341.3 ms	−0.5%
nbody	497.0 ms	494.4 ms	−0.5%
storage	689.6 ms	695.5 ms	+0.9%