perf(pd): reduce disaggregated serving latency

[sufubao]

Summary

• make detokenization and PD KV polling intervals configurable through environment variables
• reduce router-loop overhead by making metric gauge cadence configurable and avoiding uvloop busy-yields for sub-millisecond intervals

The request shared-memory handle cleanup was extracted into #1482.

Tests

• pre-commit run black --files
• pre-commit run flake8 --files