Resolve truncated normal issues

[mrTsjolder]

Attempt to implement the fix for the issue with truncated normal as discussed in #8048

[fchollet]

Thanks for the PR.

This would be a breaking behavior change, so we won't merge it. In general it is preferable to align the behavior of backend functions on the corresponding TF functions. If the docstring is misleading, we can fix the docstring.

What we can do, is merge your fix for Theano (since a different behavior for Theano would be a bug), and add a unit test checking that the distributions are equivalent across the different backends.

[mrTsjolder]

Changing the documentation of the backend functions would be a good idea indeed (this is also how tensorflow handled this issue), because, in my eyes, the option to provide standard deviation as a parameter to a distribution, implies that the returned distribution has the given standard deviation.

In the case that keras.backend.truncated_normal is really supposed to have a lower standard deviation than the one provided as argument to the function, the correct Theano function to use, would be rng.normal(shape, mean, stddev, dtype, truncate=True).

In order for the initializers to finally work correctly, I do think it is important to fix keras.initializers.VarianceScaling. Currently, it does not provide distributions with the given standard deviation when using distribution='normal', breaking the entire foundations of most initializers.

See #9963 for this alternative fix