Don't ignore expected types of New.

[odersky]

A problem arises if we typecheck an expression like new C() with expected type C[U] where C is defined like this

class C[X](using X)()

In this case, we'd like to propagate U as the type instance of X before we
resolve the using clause. To do this, we have to keep the expected result
type C[X] for typechecking the function part new C. Previously, that type was wrapped
in an IgnoredProto.

The problem was detected now since a class C with just the using clause and
no empty parameter clause was previously expanded to

class C[X]()(using X)

but is now expanded to

class C[X](using X)()

Under the previous expansion, we type checked and instantiated new C() before looking
for an argument of the using clause, so the problem did not arise.

Fixes #15664

[odersky]

It turns out an sjsJUnitTest fails without the second commit, since that one does insert an implicit conversion after a new.

This is all goes to show that what we do to deal with implicit conversions is completely crazy. It's a black art when to propagate expected types. There's often no better or worse. Almost any change we do can heal some code and break
some other code. I believe the only way to get out of this swamp is to get rid of unrestricted implicit conversions.

[prolativ]

@odersky one project in the community build is failing but this seems unrelated to your changes. I think rebasing might fix the problem