Make key-offset shifts deterministic

[ezfoom]

What changed

• Clone the source slice before each in-place key-offset shift in train_gpt.py and train_gpt_medium.py.
• Keep the intended one-token shift unchanged while removing source/destination aliasing.

Root cause

The assignments write k[:, 1:] while reading the overlapping k[:, :-1] view on CUDA. The copy kernel can observe values that another thread has already overwritten, so the result depends on execution order instead of representing a stable one-token shift.

At the record-84 tensor shape on an H100, eight repeated runs from the same input produced a maximum element delta of 7.28125 with the overlapping assignment. Cloning the RHS produced 0.0 repeat delta.

Impact

Key-offset attention becomes deterministic and matches the shift described by the surrounding comment. This affects the long-window key-offset paths in both training scripts.

Validation

• H100 overlap reproduction at shape [1, 16384, 6, 128]: overlap repeat delta 7.28125; cloned RHS 0.0
• python3 -m py_compile train_gpt.py train_gpt_medium.py
• git diff --check