WIP: typemap split on non-leaf TypeNames

[vtjnash]

There's still a few TODO items to get full use of this optimization, but I figured I'd give a sneak peek at some of the numbers I think I'm getting. Currently, this seems to give about a 50% speedup on loading precompiled files and makes the most significant dent to-date in #15048:

~/julia-master$ time ./julia -e 'using DataFrames'
real    0m32.365s
user    0m31.650s
sys 0m2.069s
~/julia-master$ time ./julia -e 'using DataFrames'
real    0m0.917s
user    0m0.871s
sys 0m0.143s

~/julia$ time ./julia -e 'using DataFrames'
real    0m24.753s
user    0m24.297s
sys 0m1.900s
~/julia$ time ./julia -e 'using DataFrames'
real    0m0.742s
user    0m0.696s
sys 0m0.139s

I think this should finally gets us to the realm where #15048 is caused by general slow startup issues (which may be helped by #5155) and not specific to reloading method tables :)

The longer technical explanation for what this patch does is that it creates a new hashmap level in the TypeMap object that splits on the uid of the names of concrete types. This gives us the following sorted levels:

• Type{LeafType} hash-table
• LeafType hash-table
• linear-sorted list of the stuff that goes here (the technical definition of stuff is jl_args_morespecific_typename)
• linear-sorted list of signatures shorter than offs (prevents matching varargs)
• Type{Concrete TypeName} hash-table (possibly split via jl_args_morespecific_typename from another sig)
• Concrete TypeName hash-table (also possibly split)
• linear-sorted list of everything else
• ::Any

[vtjnash]

I think this would have fixed #18331

[vtjnash]

This is old enough, I think it makes more sense to re-submit it as a new PR