remove redundant msg normalization + align env_response api

[mikasenghaas]

Description

Reduces calls to normalize_messages to minimal set:

• Once and unconditionally in init_state (this is safe, because not on hot path)
• Then for all other call sites (e.g. after get_prompt_messages or env_response) we use maybe_normalize_messages which emits a warning before normalizing the message to nudge users towards using the types directly to avoid potential performance bottlenecks
• Changed the env_response return type from Messages | str to just Messages to be consistent with the rest of the API (e.g. get_model_response, rollout, etc just operate on Messages and str -> vf.Messages conversion is an example of (undesired) conversion to custom types) and thus be handled the same way as list of dicts
• Also fixes all our envs to use those patterns

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update
• Test improvement

Testing

• All existing tests pass when running uv run pytest locally.
• New tests have been added to cover the changes

Checklist

• My code follows the style guidelines of this project as outlined in AGENTS.md
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• Any dependent changes have been merged and published

Additional Notes

---

Note

Medium Risk
Tightens the core MultiTurnEnv/Environment.get_model_response APIs to require Messages and introduces conditional normalization, which could break downstream custom envs/providers that still return raw dicts/strings or pass string prompts.

Overview
Standardizes environment and model prompting to operate on typed vf.Messages only: env_response() and get_model_response() now take/return Messages (no more str), and built-in envs (alphabet_sort, doublecheck, sentence_repeater, GymEnv) were updated to emit vf.UserMessage objects instead of raw dicts.

To cut hot-path Pydantic overhead, MultiTurnEnv now uses maybe_normalize_messages() (new helper) to only normalize when needed and logs a warning once when callers still return raw dicts/strings; supporting log_once/warning_once utilities were added, and parse_response_message() now constructs AssistantMessage directly instead of round-tripping through model_validate.

^{Written by Cursor Bugbot for commit cd87b2c. This will update automatically on new commits. Configure here.}

[willccbb]

Approving in advance, but let's auto-convert or hard-fail if get_prompt_messages is overridden to return dicts. would prefer auto-convert i think, writing dicts is often the more ergonomic way to do it, but fine either way (could surface a helper fn + mention in docstring?).

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}