[2.x] Reduce number of long-living instances to speed up loading

[adpi2]

After loading, instances of ScopedKey, Scope, Select or ProjectRef represent 25.6% of the total number of instances, and 27.3% of the heap size. The main idea of this PR is to reduce the number of duplicated instances using uniqueness caches.

Uniqueness cache on Scope

It reduces the number of Scope instances by 94%. On its own, it slows down the loading, because Scope.apply becomes more expensive and it is called too many times. This can be compensated by using temporary caches in Project.transform and delegates: Scope => Seq[Scope], to avoid calling Scope.apply to create too many short-lived instances.

It also reduces the number of Select by 92% and ProjectRef by 99.8%, which make them totally negligible.

Since we use a uniqueness cache we might as well pre-compute the hash code of each scope, which reduce the CPU time.

Uniqueness cache on ScopedKey (uncommited)

I think it is also worth mentioning that I tried a uniqueness cache on ScopedKey but it was not really better. Indeed, on one hand it reduced the number of ScopedKey by 59.6%, but on the other hand it was counterbalanced by the number of nodes in the cache itself.

Other changes

The first commit reduces the number of Init.GetValue by 48% for free.

Result

I am still using sttp to collect those metrics, on my 12 CPUs laptop, and JDK temurin 17 (G1GC).

It shows up a 20% speedup of the loading time compared to sbt 2.0.0-M2, and a 41% speedup compared to sbt 1.10.2. Those are rough estimations.

Compared to 2.0.0-M2

	Settings	Instances	Used Heap size	Load time
2.0.0-M2	212961	20,920 K	532 MB	7949 ms
This PR	212961	16,365 K	410 MB	6328 ms
	100 %	78 %	77 %	80 %

Compare to 1.10.2

	Settings	Instances	Used Heap size	Load time
1.10.2	197031	23,524 K	675 MB	10629 ms
This PR	212961	16,365 K	410 MB	6328 ms
	108 %	70 %	61 %	59 %

[eed3si9n]

@adpi2 Nice work!