make sure that nullfunc is not ellided

[vchuravy]

fixes #163

We don't use these functions by default to estimate the overhead,
but right now they give the answer of:

before = time_ns()
after = time_ns()
(after-before)/evals

Which is not a particular helpful estimate of the costs. With
a inference barrier we have a for-loop and a call to a function.

We explicitly note in https://github.com/JuliaCI/BenchmarkTools.jl/blob/master/doc/manual.md#understanding-compiler-optimizations
that compiler optimizations like that have to be accounted for by the user

This is more accurate and seems more stable, but requires a large number
of iterations to converge, probably means we should revitalize #94.

julia> BenchmarkTools.overhead_sample(10000)
2.8025

julia> BenchmarkTools.overhead_sample(10000)
6.3016

julia> BenchmarkTools.overhead_sample(10000)
5.5982

julia> BenchmarkTools.overhead_sample(10000)
6.3029

julia> BenchmarkTools.overhead_sample(10000)
6.2975

julia> BenchmarkTools.overhead_sample(10000)
5.5966

julia> BenchmarkTools.overhead_sample(100000)
5.79436

julia> BenchmarkTools.overhead_sample(1000000)
5.312702

julia> BenchmarkTools.overhead_sample(10000000)
3.9741005

julia> BenchmarkTools.overhead_sample(100000000)
1.59801717

julia> BenchmarkTools.overhead_sample(1000000000)
1.385175517

julia> BenchmarkTools.overhead_sample(10000000000)
1.3278229344

julia> BenchmarkTools.overhead_sample(100000000000)
1.37250904273

[jrevels]

TIL about Base.inferencebarrier, that's a dope little addition