ci(codspeed): pin benchmark runner to ubuntu-24.04 (#31746)

* ci(codspeed): pin benchmark runner to ubuntu-24.04

ubuntu-latest resolves to different runner images between the BASE
(main/staging) and HEAD (PR) runs, so CodSpeed reports 'Different
runtime environments detected' and emits false-positive regressions
(e.g. a -25.2% swing on test_completion_multi_turn in #31684, an MCP
auth fix with no LLM code changes). Pinning the runner to a fixed
image keeps BASE and HEAD on the same hardware so 1 ms swings on a
~3 ms benchmark stop blocking unrelated PRs.

Fixes #31738

* ci(codspeed): stop running benchmarks on litellm_internal_staging

The CodSpeed check flip-flops on internal staging and on PRs targeting
it (e.g. "+11.75% improvement" on one run, "-25.36% regression" on the
next) because the comparison flags "different runtime environments" and
the benchmarks are only 3-4 ms, so sub-millisecond runner noise swings
the result by 25-30%. Pinning the runner to ubuntu-24.04 in this PR
helps the head side, but the internal_staging base is still recorded on
the old unpinned runner, so comparisons keep flapping until the pin
merges and the base is re-baselined.

Until that settles, the red X's on internal staging make the OSS
project look unhealthy and confuse contributors, so drop the
litellm_internal_staging push and pull_request triggers and keep
CodSpeed running on main only.

---------

Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>

.github/workflows/codspeed.yml

@@ -4,11 +4,9 @@ on:
   push:
     branches:
       - main
-      - litellm_internal_staging
   pull_request:
     branches:
       - main
-      - litellm_internal_staging
   # Allow CodSpeed to trigger backtest performance analysis
   # in order to generate initial data
   workflow_dispatch:
@@ -23,7 +21,7 @@ concurrency:
 
 jobs:
   benchmarks:
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-24.04
     timeout-minutes: 15
 
     steps: