feat(vars): index upstream timing variables for FFI access

[gruceo]

We have identified a ~5% RPS performance regression in the Kong EE PR 11452

Upon further investigation, we found that a big chunk of CPU time is spent on ngx.var variables access.

We add upstream timing variables to the default indexed variables list, allowing efficient access through Kong's FFI interface. These variables track detailed timing information for upstream connections:

• upstream_start_timestamp_us
• upstream_connect_timestamp_us
• upstream_request_timestamp_us
• upstream_header_timestamp_us
• upstream_response_timestamp_us

These microsecond-precision timestamps enable detailed upstream latency analysis and troubleshooting of connection issues.

[chronolaw]

A temporary solution is using directive lua_kong_load_var_index in nginx.conf manually.

[gruceo]

After we merged this nginx patch in OSS, a test is failing consistently: https://github.com/Kong/lua-kong-nginx-module/actions/runs/13846604773/job/38746240223#step:8:42

#   Failed test 't/011-get_last_peer_connection_cached.t TEST 1: sanity - grep_error_log_out (req 0)'
#   at /usr/local/share/perl/5.34.0/Test/Nginx/Socket.pm line 1220.
# @@ -1,3 +1,2 @@
#  last_peer_connection_cached 0
#  last_peer_connection_cached 0
# -last_peer_connection_cached 1
# Looks like you failed 1 test of 2.

It seems related to the patch in the sense that the new timestamp variables we added are causing the upstream state to be reset or modified in a way that affects connection caching behavior because the test expects to see a cached connection being used. @chronolaw @samugi

[samugi]

looks to be unrelated to the nginx patch, probably related to the OpenResty bump - moving the conversation internally