## Problem

\`PrometheusLogger._increment_remaining_budget_metrics\` is called on every successful LLM API request and always fans out four async tasks via \`asyncio.gather\` — one each for API key, team, user, and org budget metrics. Each task calls an \`_assemble_*_object\` helper that performs a DB or cache lookup (\`get_team_object\`, \`get_key_object\`, \`get_user_object\`, \`get_org_object\`) to hydrate the entity record, then calls \`.set()\` on the corresponding Prometheus gauge.

When an operator uses \`prometheus_metrics_config\` to restrict which metrics are published, any metric not in the allowlist is created as a \`NoOpMetric()\` instance whose \`.set()\` is a silent no-op. However, neither \`_increment_remaining_budget_metrics\` nor any of the four per-entity helpers checks whether the target gauge is a \`NoOpMetric\` before doing the I/O. The entire DB/cache lookup pipeline runs, and every retrieved object is immediately discarded.

In deployments that deliberately exclude the budget gauges (\`litellm_remaining_team_budget_metric\`, \`litellm_remaining_api_key_budget_metric\`, \`litellm_remaining_user_budget_metric\`) from the allowlist, this generates four async DB…