Make conversions work for Float16 (and speed tests)

[timholy]

Curiously, currently we have this:

julia> using FixedPointNumbers

julia> x = UFixed8(0.3)
UFixed8(0.298)

julia> y = convert(Float16, x)
0.29803923f0

julia> typeof(y)
Float32

As @vchuravy knows well, that may soon be fixed on julia master (see JuliaLang/julia#17297), but since this package also supports julia 0.4 it seems sensible to fix it here.

I also was reminded about how awfully slow the tests are. Poking at this a bit, I discovered that changing two characters (see 06b6372) speeds up the tests by several fold. Moreover, I suspect there's more improvement possible, because this loop isn't type-stable, even on julia-0.5 (the type of fun is changing on each iteration). However, I confess I'm having trouble understanding the purpose of that loop: as far as I can tell, it only ever runs the tests for fun == /. @vchuravy, can you shed some light on this? I'm guessing you meant to run the two lines with @test for all 4 functions, skipping it for / when fy == 0, right?

[timholy]

OK, I pushed two more commits and have gotten the tests from ~10min to ~20s. @vchuravy, please check to see whether you think these changes are correct.

[vchuravy]

Thanks, the for loop was indeed a mistake. @test is really slower than @assert?

This looks good. lgtm

[timholy]

Much of the performance improvement came from other changes, but the @test->@assert change also had a pretty big impact.

This branch:

julia> @time include("runtests.jl")
# <output suppressed>
 19.473147 seconds (93.47 M allocations: 1.910 GB, 1.71% gc time)

If I drop the last commit that changed @test to @assert:

julia> @time include("runtests.jl")
# <output suppressed>
144.961808 seconds (3.16 G allocations: 143.512 GB, 11.94% gc time)

I'd say that's enough of an improvement to be worth it. If you disagree, I'm happy to drop the last commit.

[vchuravy]

No that looks worth it. Thank you for this.