fix(applications): RLRW returns the choices it simulates

[lenarddome]

Closes #100.

With generate=True, RLRW sampled a choice on every trial but did not return it, so its simulations could not be used for parameter or model recovery.

Changes

• _sessions.rlrw records the chosen arm of every trial (int64) and returns it; RLRW outputs it as response, the column it reads observed choices from, so simulated data can be fitted directly. Without generate, response is the observed choice.
• Documented generate and the new output in the RLRW docstring.
• Tests: simulated choices follow the returned policy (recomputed from the same seed) and the reward is that of the chosen arm; response equals the observed choices without generate; numba and Python simulate the same choices. The frozen per-trial reference predates response, so the reference comparisons leave it out.
• Version bumped to 0.26.0.dev2, with a changelog entry.

Every RLRW export now has one more column, response; other outputs are unchanged.

🤖 Generated with Claude Code