fix(examples): preserve geo3k response budget

[zhangdw156]

Summary

• Fix geo3k VLM multi-turn rollout budgeting so max_new_tokens is treated as response budget, not total prompt+response budget.
• When rollout_max_context_len is also set, use the tighter remaining context/response budget.
• Add a CPU-only regression test for fresh prompts, existing response tokens, and combined context/response caps.

Why

sampling_params["max_new_tokens"] comes from rollout_max_response_len. The current multi-turn example subtracts len(sample.tokens), which includes the prompt, so long prompts can prematurely truncate or reduce the model's response budget before generation starts.

Tests

• uvx --from pytest pytest tests/test_geo3k_vlm_multi_turn_budget.py -q
• uvx ruff check examples/geo3k_vlm_multi_turn/rollout.py tests/test_geo3k_vlm_multi_turn_budget.py
• git diff --check