evalv3: infinite work and OOM panic regression where evalv2 is very fast

[mvdan]

env CUE_EXPERIMENT=evalv3=0
exec cue export

env CUE_EXPERIMENT=evalv3=1
exec cue export

-- file.cue --
package p

out: #job.step & "foo"

#job: (#Workflow & {job: _}).job

#Workflow: {
	job: step: string

	#matrixConfig: [...#matrixConfig] | string
	matrix?: [string]: [...#matrixConfig]
}

As of 4e8e027, I cannot run this testscript, as it doesn't deal well with commands that never finish. Below is me running the same file.cue file with the old and new evaluator manually, with timings:

$ time CUE_EXPERIMENT=evalv3=0 cue export repro.cue; echo --; time CUE_EXPERIMENT=evalv3=1 cue export repro.cue
{
    "out": "foo"
}

real	0m0.014s
user	0m0.007s
sys	0m0.009s
--
^C
real	0m17.613s
user	0m52.158s
sys	0m3.985s

Note that I had to kill the new evaluator run at about 25GiB of memory used, so that it wouldn't start forcing my laptop to use swap memory.