fix throughput metric attribution order crash in metrics

[geocine]

• fixes metrics crashing when vllm:time_per_output_token_seconds appears before TTFT metrics by using trpt_sample
  labels instead of ttft_sample.
• prevents UnboundLocalError and ensures throughput stats are attributed to the correct model.
• scope is limited to throughput parsing; no behavior changes elsewhere.
• run metrics with a Prometheus payload where vllm:time_per_output_token_seconds is emitted before
  vllm:time_to_first_token_seconds -> UnboundLocalError at metrics.py:66.

Repro (no vLLM Required)

from prometheus_client.parser import text_string_to_metric_families
import util.mlib as mlib
from src.metrics.metrics import metrics

RAW = """
# HELP vllm:time_per_output_token_seconds summary
# TYPE vllm:time_per_output_token_seconds summary
vllm:time_per_output_token_seconds_sum{model_name="m1"} 2
vllm:time_per_output_token_seconds_count{model_name="m1"} 1
# HELP vllm:time_to_first_token_seconds summary
# TYPE vllm:time_to_first_token_seconds summary
vllm:time_to_first_token_seconds_sum{model_name="m1"} 1
vllm:time_to_first_token_seconds_count{model_name="m1"} 1
"""

mlib.get_metrics = lambda: list(text_string_to_metric_families(RAW))
mlib.get_running_models = lambda: []
mlib.get_cuda_gpu_infos = lambda: []

metrics(None)

Before fix: UnboundLocalError at metrics.py:66 (throughput metrics processed before TTFT).
After fix: table renders without error and throughput shows for model m1.