This way, we have only one implementation for all the versions of
Scala. This will become more useful as we make its implementation
less trivial.

We also rename it back to `UniquenessCache`, which is the name it
has on the JVM, and reinstate the `Null` lower bound on `V`. There
was no reason to name it differently, and keeping the name may be
better for binary compatibility (even though it is `private[scala]`
and nobody should be using it). It is still different for the key
type, which is fixed to `String`.

They were changed in 757949e to
use a `js.WrappedArray` by default, instead of an `ArrayBuffer`.

Unfortunately, if we want to make the scalalib independent from the
Scala.js library, this is not something we can do anymore.

Ideally, we would override `ArrayBuffer` itself with an
implementation based on `js.Array`. We cannot do that either
because its `array` field is protected, and it is not final.

It has specific overrides in `overrides-2.12` and `overrides-2.13`.

…obal`.

This is required not to depend on the Scala.js library from
`ExecutionContext`.

This removes references to run-time components of
`js.WrappedDictionary` that the IR cleaner cannot get rid of.

We use it in the scalalib, so that we do not generate references
to `toScalaVarArgs`, which uses custom collections from the
Scala.js library.

It is configured differently than for the javalib. We allow
references to `scala.*` (obviously) but disallow references to
`scala.scalajs.*`. We also avoid rewriting Scala functions to JS
functions, since that changes the binary APIs.

We use a local copy of `sjs.runtime.AnonFunctionN` classes in
`scala.runtime.*` to support the codegen of lambdas.