produce LineInfoNode directly in lowering

[JeffBezanson]

This makes the output of lowering closer to what is actually needed, cutting down on extra conversion work.

For top-level code and (upcoming) OpaqueClosure, the extra step won't be needed any more, and for methods all that's needed is to insert the correct name in the location info.

[simeonschaub]

This change seems to have regressed stacktrace printing for inlined functions, which is quite important e.g. for debugging Cassette. Compare this PR:

julia> using Cassette

julia> Cassette.@context NaNCtx;

julia> Cassette.posthook(::NaNCtx, out::Number, f, args...) = isnan(out) ? error("$f$args returned $out") : nothing

julia> checked(f) = Cassette.overdub(NaNCtx(), f)
checked (generic function with 1 method)

julia> checked() do
           0 / 0
       end
ERROR: div_float(0.0, 0.0) returned NaN
 Stacktrace:
  [1] error(s::String)
    @ Base ./error.jl:33
  [2] posthook
    @ ~/Documents/julia/bisect.jl:100 [inlined]
  [3] overdub
    @ ./float.jl:407 [inlined]
  [4] overdub
    @ ./int.jl:93 [inlined]
  [5] overdub(::Cassette.Context{nametype(NaNCtx), Nothing, Nothing, Cassette.var"##PassType#251", Nothing, Nothing}, ::typeof(/), ::Int64, ::Int64)
[...]

with 1.5, where 3 and 4 correctly show up as calls to /:

Stacktrace:
 [1] error(::String) at ./error.jl:33
 [2] posthook at ./REPL[6]:1 [inlined]
 [3] / at ./float.jl:407 [inlined]
 [4] / at ./int.jl:92 [inlined]
 [5] overdub(::Cassette.Context{nametype(NaNCtx),Nothing,Nothing,Cassette.var"##PassType#255",Nothing,Nothing}, ::typeof(/), ::Int64, ::Int64) at /home/simeon/.julia/packages/Cassette/Wjztv/src/overdub.jl:0

Is this something that just needs a fix, or would we need to revert this PR?