more precise eltype for some Generator subtypes

[nsajko]

Fixes #41519

Fixes #48448

[mcabbott]

TIL that x for x in A is noticed by the parser, and produces this type:

julia> (x for x in 1:3)
Base.Generator{UnitRange{Int64}, typeof(identity)}(identity, 1:3)

julia> (identity(x) for x in 1:3)
Base.Generator{UnitRange{Int64}, var"#13#14"}(var"#13#14"(), 1:3)

(Was true on Julia 1.6, not true on 1.0.)

I wonder what would break if this instead returned A, i.e. we made this (x for x in 1:3) === 1:3.

[nsajko]

Yeah, I discovered the same thing while working on the PR, and wonder the same thing, too.

[nsajko]

Actually, I know what would break. map(f, x) is implemented as something like collect(Iterators.map(f, x)). This relies on the fact that Generator has EltypeUnknown.

[nsajko]

Lots of code assumes that, e.g., map(identity, x) will return a collection with more precise eltype than x has. This is what relies on Generator having EltypeUnknown.