Regression: symbols include corrupted duplicate constructor symbol for attribute

[auduchinok]

[<Measure>] type m

Expected (previous results):

Actual:

It includes an extra symbol with incorrect name and corrupted range.

It seems it comes from #19252.

[auduchinok]

Ideally, only the MeasureAttribute constructor should be returned, i.e. the type symbol should be overwritten in the sink, but this is easy enough to workaround, unlike the new erroneous symbol.

[auduchinok]

This also breaks cases where code is rewritten during analysis:

type T() =
    static member M1(?a) = ()

Ideally, the rewritten symbols should have synthetic ranges.

[auduchinok]

OK, it's even worse and happens not only for attributes. Each constructor usage records a broken extra symbol with an incorrect range.
In addition to the wrong name, the duplicate symbol also loses the type argument information in cases like this:

type Base<'T>() =
    abstract M1: 'T -> unit
    abstract M2: 'T -> unit

{ new Base<int>() with
    override this.M1 _ = () }

The expected constructor symbol should look like this (but it's effectively overwritten by the broken one):