QuadraticTimeMMD.compute_variance_h0 fails on aarch64

[vigsterkr]

aarch64's float operations seems to be not so accurate :)
in QuadraticTimeMMD there's a 4.76837158203125e-07 error instead of 1E-10.
would it be a problem that we lower an error a bit:

[ RUN      ] QuadraticTimeMMD.compute_variance_h0
/home/buildslave/xenial_-_libshogun_-_aarch64/build/tests/unit/statistical_testing/QuadraticTimeMMD_unittest.cc:277: Failure
The difference between var and 0.0042963027954101562 is 4.76837158203125e-07, which exceeds 1E-10, where
var evaluates to 0.0042967796325683594,
0.0042963027954101562 evaluates to 0.0042963027954101562, and
1E-10 evaluates to 1e-10.

[lambday]

Hi @vigsterkr. I think it should be okay. These tests use a precomputed kernel matrix, which is of float32_t type. I think FLT_EPSILON should be okay in this case?

[karlnapf]

@lambday want to fix? :) We need your help ...

[lambday]

@karlnapf sure! Will send a patch today.. :)

[vigsterkr]

The difference between var and 0.0042963027954101562 is 4.76837158203125e-07, which exceeds 1.19209289550781250000000000000000000e-7F, where
var evaluates to 0.0042967796325683594,
0.0042963027954101562 evaluates to 0.0042963027954101562, and
1.19209289550781250000000000000000000e-7F evaluates to 1.1920928955078125e-07.``` :)