[change] Remove concatenate in discrete action probabilities to improve inference performance

[ervteng]

Proposed change(s)

Discrete action models in ML-Agents produces outputs for action probabilities, one per branch, concatenates them, then breaks them out again to get the actions themselves. see:

This causes redundant cache-copying behavior in Barracuda and may slow inference performance quite a bit.

This PR changes the network architecture to below to avoid this behavior:

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

[chriselion]

Maybe add something to the changelog?