docs: replace vLLM metric names that current vLLM no longer exports - #2835

Three places point readers at vLLM metrics that a current vLLM server
does not expose, so the queries return no data:

- guides/turn-priority-fairness: vllm:gpu_prefix_cache_hit_rate was a
  V0-engine gauge. The V1 engine (and the vLLM 0.26 this guide was
  validated on) only exports the vllm:prefix_cache_hits_total and
  vllm:prefix_cache_queries_total counters, so show the hit ratio as a
  rate of the two.
- docs/architecture/autoscaling/wva.md: the avg_itl rows name
  vllm:time_per_output_token_seconds, but the WVA collector (v0.9.0,
  internal/collector/registration/queueing_model.go) queries
  vllm:inter_token_latency_seconds. vLLM renamed the histogram in 0.11
  and removed the old name in 0.15.
- the DigitalOcean monitoring dashboard: the "Time Per Output Token
  Latency" panel queries vllm:time_per_output_token_seconds_bucket;
  switch it to vllm:inter_token_latency_seconds_bucket.

Signed-off-by: YoungjuKim <[email protected]>