perf: reduce buffered transform copies in node streams
Context:
- Avoid repeated Buffer.concat in createBufferedTransformNode by buffering chunks and flushing once.
- Keep callback timing unchanged while reducing per-chunk allocation overhead.
Benchmark (micro scenario, 300 iterations, 30 warmup, parent-file swap):
- createBufferedTransformNode only median: 0.070ms -> 0.052ms (+26.85% faster)
- Node continueDynamicHTMLResume median: 0.657ms -> 0.571ms (+13.01% faster)
- Node continueStaticPrerender median: 0.609ms -> 0.611ms (-0.40%)
