Reduce required accuracy of some QuadGK integrals

This regressed in v2.8.1:
<#134 (comment)>.

test/runtests.jl

@@ -848,10 +848,10 @@ end
 
 @testset "QuadGK" begin
     @test QuadGK.quadgk(cos, x, y)[1] ≈ sin(y) - sin(x)
-    @test QuadGK.quadgk(sin, -y, y)[1] ≈ cos(-y) - cos(y) atol = eps(Float64)
+    @test QuadGK.quadgk(sin, -y, y)[1] ≈ cos(-y) - cos(y) atol = 2 * eps(Float64)
     @test QuadGK.quadgk(exp, 0.4, x)[1] ≈ exp(x) - exp(0.4)
     @test QuadGK.quadgk(sin, w, 2.7)[1] ≈ cos(w) - cos(2.7)
-    @test QuadGK.quadgk(t -> cos(x - t), 0, 2pi)[1] ≈ measurement(0) atol = 7e-16
+    @test QuadGK.quadgk(t -> cos(x - t), 0, 2pi)[1] ≈ measurement(0) atol = 2e-13
     @test QuadGK.quadgk(t -> exp(t / w), 0, 1)[1] ≈ w * (exp(1 / w) - one(x))
     for a in (w, x, y)
         @test QuadGK.quadgk(t -> 1 / abs2(t / a), 1, Inf)[1] ≈ a ^ 2
@@ -866,7 +866,7 @@ end
     @test QuadGK.quadgk(cos, x, y)[1] ≈
         @uncertain(((x,y) -> QuadGK.quadgk(cos, x, y)[1])(x, y))
     @test QuadGK.quadgk(sin, -y, y)[1] ≈
-        @uncertain((x -> QuadGK.quadgk(sin, -x, x)[1])(y)) atol = 1e-10
+        @uncertain((x -> QuadGK.quadgk(sin, -x, x)[1])(y)) atol = 5e-10
     @test QuadGK.quadgk(exp, 0.4, x)[1] ≈
         @uncertain((x -> QuadGK.quadgk(exp, 0.4, x)[1])(x))
 end