no hacky overwriting of typeinf_local

[aviatesk]

And the changes also remove the dirty overwriting of typeinf_local

[aviatesk]

okay, it turns out that enabling optimization (and thus integrating error reports with internal code cache) needs some work.
I would like to make this PR only remove the hacky overwriting of typeinf_local. Getting rid of that has one drawback that now we can't get "as much errors as possible" in a single gf call (e.g. fib("1000") in demo.jl) because native interpreter's typeinf_local early escape as soon as it uncovers Union{}. So for example now we only get 4 error reports for demo.jl, but it's acceptable imho because at least one error point would be reported for one toplevel gf call anyway. Apart from that, it is clearly better and preferable in terms of maintenance and performance.