[codex] Truncate inference rank logs with tee

[samsja]

Summary

Replaces the startup rm -f logs/inference/*.log cleanup with writer-side truncation for rank-specific inference logs.

Why

The node-level inference log already matches trainer/orchestrator behavior: it gets a startup line through truncating tee, then appends server output during the same run. The auxiliary node_*_rank*.log files did not have that initial truncating write and only used tee -a, so they could retain content when reusing an output directory.

Impact

• Removes the explicit inference log deletion added in [codex] Truncate inference rank logs on launch #2814.
• Keeps node_*.log header-plus-append behavior.
• Uses truncating tee for node_*_rank*.log outputs.

Validation

• git diff --check -- src/prime_rl/templates/_launch_rank.sh.j2 src/prime_rl/templates/inference.sbatch.j2 src/prime_rl/templates/multi_node_rl.sbatch.j2
• Loaded the edited Jinja templates with jinja2.Environment(...).get_template(...)