http: use intrusive lists in ConnectionsList

[mcollina]

Every HTTP message was performing multiple erase and insert operations on the two std::set instances in ConnectionsList, which are ordered by a key (last_message_start_) that mutates on every message. In a hello-world server profile this showed up as ~2.2% of total CPU cycles spent in red-black tree rebalancing plus tree-node allocations, all on the per-request hot path.

This PR replaces both sets with intrusive doubly-linked lists:

• A node is linked to itself when it is not in a list, so Remove() is idempotent and the destructor unlinks automatically, which also removes the old "Pop from the lists BEFORE resetting last_message_start_" footgun.
• Membership in the "all connections" list no longer changes per message, so the Pop+Push pair in on_message_begin/on_message_complete is gone entirely.
• Updating the active list is now a single O(1) unlink + tail append with no allocations.
• Appending to the tail keeps the active list ordered by last_message_start_ because uv_hrtime() is monotonic, so expired() semantics are unchanged. The arrays returned by all()/idle()/active() are now in insertion order instead of comparator order; none of their consumers depend on ordering.

Benchmark results (Linux, i7-7700, server and benchmarker pinned to separate cores):

                                                                                           confidence improvement accuracy (*)   (**)  (***)
http/simple.js duration=5 chunkedEnc=0 c=50 chunks=1 len=4 type='bytes' benchmarker='wrk'          **      4.06 %       ±2.41% ±3.23% ±4.25%
http/simple.js duration=5 chunkedEnc=0 c=500 chunks=1 len=4 type='bytes' benchmarker='wrk'                 2.35 %       ±2.45% ±3.29% ±4.34%

A separate interleaved wrk A/B (hello-world server, -t2 -c50, 5 alternating pairs) shows +4.9% (26,599 → 27,914 req/s), with the patched binary winning every pair. A perf profile of the patched binary confirms the _Rb_tree symbols are gone and malloc traffic in the parser path is roughly halved.

This PR was prepared with the help of AI assistance.

[nodejs-github-bot]

Review requested:

• @nodejs/http
• @nodejs/net

[codecov]

Codecov Report

❌ Patch coverage is 91.11111% with 4 lines in your changes missing coverage. Please review.
✅ Project coverage is 90.31%. Comparing base (347e266) to head (41044a2).
⚠️ Report is 121 commits behind head on main.

Files with missing lines	Patch %	Lines
src/node_http_parser.cc	91.11%	4 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff            @@
##             main   #65296     +/-   ##
=========================================
  Coverage   90.30%   90.31%             
=========================================
  Files         759      751      -8     
  Lines      248294   250241   +1947     
  Branches    46860    47294    +434     
=========================================
+ Hits       224220   226002   +1782     
- Misses      15516    15618    +102     
- Partials     8558     8621     +63     

Files with missing lines	Coverage Δ
src/node_http_parser.cc	84.72% <91.11%> (-0.26%)	⬇️

... and 145 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[ronag]

Have you benchmarked against an unordered_set or just a plain vector?

I suspect those would be more cache friendly and avoid pointer chasing.