Avoid PGRAPH register usage analysis in tracing code

[JayFoxRox]

The actual tracing code in nv2a-trace should know as little as possible about the GPU. All decoding and analysis tools should be separate or be integrated from separate code using some interface (like callbacks).

However, the existing code did a diff between PGRAPH register dumps as part of the command processor. This breaks this design goal.

This PR modifies it so the state is only written to disk. A new tool to compare register dumps can be created later, based on the removed code (or written from scratch).

Similar changes are planned for the texture and surface dumping in the future.