Numerical issue on OSX

[JohanMabille]

The lgamma test with negative inputs have been disabled for architecture supporting AVX but not AVX2, because they lead to small differences for a few inputs:

idx = 1533 - input = -3.9555075 - expected = 0.00443037619386222 - result = 0.00443037619386457
idx = 1539 - input = -3.9553725 - expected = 0.00162255524386734 - result = 0.00162255524386876
idx = 1540 - input = -3.95535 - expected = 0.00115548170832613 - result = 0.00115548170832745
idx = 1548 - input = -3.95517 - expected = -0.00257193023599123 - result = -0.00257193023598964
idx = 1549 - input = -3.9551475 - expected = -0.00303671350222201 - result = -0.00303671350222045
idx = 1550 - input = -3.955125 - expected = -0.00350124356247306 - result = -0.00350124356247113
idx = 1551 - input = -3.9551025 - expected = -0.00396552066902212 - result = -0.00396552066901967

However, the issue seems due to the platform (OSX) rather that the suported instrution set. Indeed, building on Linux with avx2 explicitly disabled results in passing tests (i.e., not nuemrical differences).