eliminate stage0 inference

[vtjnash]

all it did was ensure that everything in specializations was inferred
and require a lot of complex infrastructure to make it acceptable to
replace a module during compilation

but instead we can just infer everything in specializations directly

(cherry-picked from #15918)

[JeffBezanson]

I see possible regressions here.

On commit 0810634

     * linalg/triangular     in 444.61 seconds, maxrss  861.92 MB
     * numbers               in  96.93 seconds, maxrss  338.26 MB

and on master 9d4d36c

     * linalg/triangular     in 471.98 seconds, maxrss  908.10 MB
     * numbers               in 130.09 seconds, maxrss  351.44 MB

[vtjnash]

if i had to guess, it's that the leaftype kind ::DataType gets added to a cache somewhere, and that prevents the runtime from doing more (better) specialization on it later

but my bisect indicates it is caused by #15934, not this one

[JeffBezanson]

That's interesting. I don't see how #15934 can add time to anything except building a system image, unless the mere presence of extra functions slows things down somehow. But compared to how much stuff we compile during the tests the extra code doesn't seem too significant.

[vtjnash]

my thought was that we compiled something for f(::DataType) instead of for each f(::Type{T}) and that impacted the cache. alternatively, perhaps simply the presence of a few extra cache entries somewhere critical made lookup slower for some critical item? there's still a lot of optimization opportunities for adding short-circuit paths out of the typemap lookups

[tkelman]

contrib/build_sysimg.jl needs fixing