Add frozen transformer next-token probes for cycle 5

[Al-does]

Summary

• restore all variant-1 cycle-5 checkpoints from their pre-rename B2 artifact prefix, including a compatibility shim for the legacy task path
• restore only the local RLModule for inference, avoiding the saved distributed training topology and its Ray worker resource demands
• train stop-gradient transformer probes at context lengths 1 and 10, with and without the selected action
• optimize exact next-token soft cross-entropy and report forward KL, sampled-token metrics, and joins to the existing affine belief MSE
• add cross-seed correlation/plot tooling and focused unit coverage, including the optimizer loop

Test plan

• uv run pytest -q tests/test_cycle5_next_token_probe.py
• cycle-5 next-token probe smoke run against the B2 initial checkpoint
• uv run pytest -q -m "not slow"