Fix #7149: Revert inference change

[odersky]

With #7114, we had a new failure scenario: A pair of <indent> ... <outdent> created a block where none was before which caused an additional interpolation step in Typer. The interpolation would instantiate type variables that occurred only co- or contra-variantly in the result type. This caused a problem for inferring type arguments for asInstanceOf. An example is in type-propagation.scala:

def foo: String =
    "abc".asInstanceOf

Significant indentation expanded this fragment as if it was

def foo: String = {
    "abc".asInstanceOf
}

The interpolation then inferred the result type of asInstanceOf to be Nothing, even though a result type String was known.

The original fix in #7130 was to not interpolate to Nothing in this case. Unfortunately, that broke type inference. #7149 is one example, the example in fastparse another.

One could argue that asInstanceOf should always get explicit type arguments, since inference from result type is very fragile. But the fact is there's a lot of code that uses asInstanceOf in the way shown above.

What this PR does instead is not to create a Block around a single indented expression. So, we generate the same trees as before.

[liufengyun]

LGTM