perf(grpo): use JSONL append-only format for training log

[abrichr]

Problem

trainer.py:674-677 rewrites the entire JSON training log on every step. With 1000 steps, the file grows continuously and the O(N) full rewrite becomes slow.

Proposed Fix

Switch to JSONL (one JSON line per entry, append-only) which is O(1) per write and easier to process with streaming tools.

[abrichr]

Closing: resolved by replacing the custom GRPO trainer with a minimal TRL bridge (see PR #34). The custom JSON training log (full O(N) rewrite per step) is eliminated. TRL integrates natively with WandB and TensorBoard for training metrics logging, which are both append-only and streaming-friendly.