[bug-fix] Add clipping to PyTorch policy, fix initialization

[ervteng]

Proposed change(s)

This PR is motivated by the perpetually lower resulting rewards that Torch PPO gets on ragdoll environments, especially Walker. It addresses two differences between Torch and TF PPO:

1. Torch layers were initialized using Kaiming Normal, but PyTorch by default assumes that you're using Leaky ReLU and adds a default gain of sqrt(2). This PR changes the nonlinearity assumed in the initializer to linear, removing that extra factor of sqrt(2). This initialization was the main cause of wildly higher value losses in Torch.
2. In TensorFlow, when not using tanh squashing on Gaussian outputs (PPO), we truncate the output by clip(action, -3, 3) / 3. This PR adds the same thing to Torch.

Types of change(s)

• Bug fix
• New feature
• Code refactor
• Breaking change
• Documentation update
• Other (please describe)

Checklist

• Added tests that prove my fix is effective or that my feature works
• Updated the changelog (if applicable)
• Updated the documentation (if applicable)
• Updated the migration guide (if applicable)

Other comments

[vincentpierre]

Don't forget to cherry pick to release_10_branch after merge.