Use roundeven instead of rint for round(x)

[vchuravy]

The llvm.rint intrinsic is sensitive to the floating-point
environment, so determined enough users can change the definition
of round(x, RoundNearest) to not be rount to even.

julia> round(2.5, RoundNearest)
2.0
julia> Base.Rounding.setrounding_raw(Float64, Base.Rounding.to_fenv(RoundUp))
julia> round(2.5, RoundNearest)
3.0
julia> roundeven(x::Float64) = ccall("llvm.roundeven.f64", llvmcall, Float64, (Float64,), x)
round_even (generic function with 1 method)
julia> roundeven(2.5)
2.0

The intrinsic was introduced with LLVM 11, and roundeven will be part
of the C23 standard (problematic for us due to runtime_intrinsics.cpp).

Sadly performance isn't looking to hot. Using @simonbyrne benchmark from
#8750 (comment)

julia> roundeven_llvm(x::Float64) = ccall("llvm.roundeven.f64", llvmcall, Float64, (Float64,), x)
roundeven_llvm (generic function with 1 method)

julia> rint_llvm(x::Float64) = ccall("llvm.rint.f64", llvmcall, Float64, (Float64,), x)
rint_llvm (generic function with 1 method)

julia> function test_roundeven_llvm(N)
           t = 1.0
           s = 0.0
           for i = 1:N
               t += eps()
               s += roundeven_llvm(t)
           end
           s
       end
test_roundeven_llvm (generic function with 1 method)

julia> function test_rint_llvm(N)
           t = 1.0
           s = 0.0
           for i = 1:N
               t += eps()
               s += rint_llvm(t)
           end
           s
       end
test_rint_llvm (generic function with 1 method)

julia> @time test_roundeven_llvm(100_000_000);
  0.274178 seconds

julia> @time test_roundeven_llvm(100_000_000);
  0.274118 seconds

julia> @time test_rint_llvm(100_000_000);
  0.087601 seconds

julia> @time test_rint_llvm(100_000_000);
  0.087860 seconds

On x86:

julia> @code_native roundeven_llvm(2.5)
	.text
	movabsq	$roundevenf64, %rax
	jmpq	*%rax
	nopl	(%rax)

e.g it's implemented with a function call to roundevenf64, which is implemented by libc.

[mkitti]

A custom x86 with sse4.1 lowering for llvm.roundeven was added with LLVM 12
https://reviews.llvm.org/D84592
https://reviews.llvm.org/rGc4823b24a41b9e3552af7781e05ed1784fb84f46

[ViralBShah]

@simonbyrne your thoughts here?

[simonbyrne]

Since we're now using a newer LLVM, this should be fine, and we should probably make the change.
I tried rebasing, but not sure how to deal with runtime intrinsics: @vchuravy?