Document CUDA training equivalence and cleanup

[ferdinandobons]

What changed

• document the measured quality equivalence between separate-head and fused QKV attention
• record the requirements for an unambiguous clean training start
• record the removal of the 14 obsolete experiment checkpoints and the retained production checkpoint policy

Why

The CUDA training guide should contain the final verified configuration and clearly distinguish a one-time cleanup from normal checkpoint retention and resume behavior.

Impact

This is a documentation-only update. It does not modify the model, training loop, dataset, or checkpoint format.

Validation

• python -B tools/validate_learngpt.py
• git diff --check