randn results in inf values in large arrays

[coquelin77]

Description
see title

To Reproduce

    features = ht.random.randn(100000, 320, split=0)  # .astype(T)
    print(torch.nonzero(torch.isinf(features.larray)))

[1] tensor([[17552,   116],
[1]         [32690,   262],
[1]         [36752,    84],
[1]         [43018,   240],
[1]         [45453,   103]])
[0] tensor([[11522,   258],
[0]         [29520,   280],
[0]         [32970,   185],
[0]         [37122,    75]])

Expected behavior
no inf values

Version Info
current master