Fix numerical instability of A matrix

[yesung31]

SIGReg was sometimes returning NaN. This fixes the divided by 0 issue.

Reproducible by the following script on MacBook Air M2.

import lejepa
import torch

torch.set_default_device("mps")
torch.set_default_dtype(torch.bfloat16)

rng = torch.Generator("mps").manual_seed(42)
a = torch.randn((2, 64, 1), generator=rng)

SIGReg = lejepa.multivariate.SlicingUnivariateTest(
    univariate_test=lejepa.univariate.EppsPulley(), num_slices=1024
)
SIGReg.global_step.fill_(5433)

loss = SIGReg(a)
print(SIGReg.global_step, loss)