perf(spend_tracking): index LiteLLM_SpendLogs by (api_key, startTime)

Every per-key spend read filters WHERE api_key = ... AND startTime in a
range (/spend/logs?api_key=..., the key-filtered UI logs page, external
billing readers), but LiteLLM_SpendLogs carries no api_key index, so each
such query scans every logged request on the instance. Composite with
startTime to match the query shape, mirroring the existing
(startTime, request_id) composite.

Measured on Postgres 18 with 1.6M spend rows: one key's 7-day SUM goes
from a 93.9ms parallel seq scan to a 0.7ms bitmap index scan (~140x);
a batch job reading per-key spend for 10k keys went from 20.5s to 1.7s.

litellm-proxy-extras/litellm_proxy_extras/migrations/20260823000000_add_spend_logs_api_key_starttime_index/migration.sql

@@ -0,0 +1,2 @@
+-- CreateIndex
+CREATE INDEX IF NOT EXISTS "LiteLLM_SpendLogs_api_key_startTime_idx" ON "LiteLLM_SpendLogs"("api_key", "startTime");

litellm-proxy-extras/litellm_proxy_extras/schema.prisma

@@ -647,6 +647,7 @@ model LiteLLM_SpendLogs {
   @@index([startTime, request_id])
   @@index([end_user])
   @@index([session_id])
+  @@index([api_key, startTime])
 }
 
 // View spend, model, api_key per request

litellm/proxy/schema.prisma

@@ -647,6 +647,7 @@ model LiteLLM_SpendLogs {
   @@index([startTime, request_id])
   @@index([end_user])
   @@index([session_id])
+  @@index([api_key, startTime])
 }
 
 // View spend, model, api_key per request

schema.prisma

@@ -647,6 +647,7 @@ model LiteLLM_SpendLogs {
   @@index([startTime, request_id])
   @@index([end_user])
   @@index([session_id])
+  @@index([api_key, startTime])
 }
 
 // View spend, model, api_key per request