[WIP] Update expected attention mean following precision changes in sigmoid

[guillaumekln]

Fixes #781.

[guillaumekln]

I was focused on GPU tests but the expected mean on CPU did not change. I'm looking for how to test this effectively on both CPU and GPU.

[qlzh727]

Thanks for taking care of this issue while I am away.

If the current issue is that GPU and CPU has different precision, you probably can split the test with

if test_utils.is_gpu_available():
....

[guillaumekln]

See #1085.