fix: max_steps=0 falsy fallback and planning remaining_steps override

[ErenAta16]

Summary

Fixes #2458 and #2510, two related bugs in how max_steps propagates through a run.

#2458: max_steps=0 silently ignored

run() does max_steps = max_steps or self.max_steps. Since max_steps: int | None = None and 0 is falsy, run(task, max_steps=0) silently falls back to the agent's constructor default instead of running zero steps.

Fixing this exposed a pre-existing edge case: when the loop body never executes (zero steps), action_step was never assigned, but the max-steps-reached branch unconditionally did yield action_step, raising UnboundLocalError. Initialized action_step = None before the loop and guarded the yield.

#2510: planning remaining_steps placeholder never interpolated, and uses the wrong budget

Two stacked issues on the same value:

1. The three built-in planning templates (code_agent.yaml, structured_code_agent.yaml, toolcalling_agent.yaml) use {remaining_steps} (single brace). populate_template renders with Jinja2 (Template(..., undefined=StrictUndefined)), which requires {{ }} for interpolation, single braces are literal text. So the model receives the literal string {remaining_steps} instead of a number.
2. Even fixed, the value was computed from self.max_steps (the constructor default) inside _generate_planning_step, not the effective per-run budget. When run(task, max_steps=N) overrides the default, the execution loop correctly uses N, but planning was still reading the constructor value. Threaded max_steps through _run_stream → _generate_planning_step so both agree.

Testing

• test_run_max_steps_zero_is_respected: run(..., max_steps=0) takes zero action steps and returns cleanly (no UnboundLocalError).
• test_planning_step_remaining_steps_uses_per_run_max_steps_override: planning text contains the per-run override's remaining-steps number, not the constructor default's, and never contains the literal {remaining_steps} placeholder.
• Updated test_planning_step's call site for the new _generate_planning_step(..., max_steps=...) signature.

Ran the full tests/test_agents.py suite before and after against unfixed code to isolate what's actually caused by this change: the same ~15 failures are present identically on unfixed main (missing numpy/transformers extras, from_folder version-compat fixtures, one ImportError, one flaky timing assertion, all unrelated to max_steps), confirming this diff doesn't introduce any new failures.