stop propagating the dict to precompile workers via repr

[KristofferC]

I completely misjudged the cost of repr on a heavily nested dict:

julia> using BenchmarkTools

julia> Base.parsed_toml("/home/kc/.julia/environments/v1.6/Manifest.toml");

julia> @btime repr(d);
  112.659 ms (32433 allocations: 7.35 MiB)

waaat(!!??).

With that time for repr, we just have to stop propagating the dict to the workers via this method. At least the workers will just parse the toml dict once each.

Noticed by @ianshmean when his parallel precompile PR in Pkg started lagging.

The culprit seems to be

  155 @Base/show.jl:763; show(io::IOBuffer, x::Type)
  ╎    ╎    ╎    ╎    ╎    ╎    155 @Base/show.jl:637; show_typealias(io::IOBuffer, x::Type)
  ╎    ╎    ╎    ╎    ╎    ╎     1   @Base/show.jl:546; make_typealias(x::Type)
 1╎    ╎    ╎    ╎    ╎    ╎    ╎ 1   @Base/reflection.jl:801; uniontypes(x::Any)
  ╎    ╎    ╎    ╎    ╎    ╎     64  @Base/show.jl:551; make_typealias(x::Type)
  ╎    ╎    ╎    ╎    ╎    ╎    ╎ 64  @Base/reflection.jl:98; names
29╎    ╎    ╎    ╎    ╎    ╎    ╎  64  @Base/reflection.jl:98; #names#10
  ╎    ╎    ╎    ╎    ╎    ╎    ╎   35  @Base/sort.jl:735; sort!
  ╎    ╎    ╎    ╎    ╎    ╎    ╎    35  @Base/sort.jl:747; #sort!#8

which I'm guessing is the new fancy type alias printing.

We could perhaps go via TOML.print:

julia> @btime sprint(TOML.print, d);
  2.135 ms (20723 allocations: 994.95 KiB)

or serialization

julia> @btime serialize(IOBuffer(), d)
  409.462 μs (2913 allocations: 200.86 KiB)

but that can be a follow-up (neither of these are available in Base so it isn't trivial to swap).

Edit: Fun fact, names gets called 308 times for this repr.

[KristofferC]

I'll merge this kinda quickly because it's so bad.