Replace torch-scatter with custom ragged reduce operation in Triton

[cswinter]

The torch-scatter dependency is a huge PITA.
We currently use it for reducing across sequences in a ragged batch to get inputs to the value head:

https://github.com/entity-neural-network/incubator/blob/85cd666f3401ca0d9eebfd0b6603e14de2311b4a/rogue_net/rogue_net/actor.py#L110-L112

This is not a very complex operation, we should just replace it with a custom Triton op and yeet the torch-scatter dependency.

[cswinter]

Before implementing the triton kernel, we should add a pure torch version, which we'll need anyway when running on CPU and as a reference to test the triton kernel: #2