fix(http): release the cached addresses of keep-alive sessions in process mode

[EJ-Research]

Fixes #6283

Bug

• A keep-alive HTTP request from a non-loopback address caches its server_addr and remote_addr strings per session in the worker
• [6.2]add server_addr #5994 in v6.2.0 extended this cache from SWOOLE_BASE to SWOOLE_PROCESS with dispatch_mode 2
• php_swoole_http_server_onClose() erases the entries but on_before_start() still installs it in SWOOLE_BASE mode only
• So in SWOOLE_PROCESS mode every such connection keeps two strings in the worker heap until the worker exits
• A user onClose callback does not help since php_swoole_server_onClose() does not touch the cache

One worker and the growth of memory_get_usage() in the worker per connection

Case	Before	After
HTTP/1.1 keep-alive closed by the client	80.00 bytes	0
The same with empty onConnect and onClose callbacks	80.00 bytes	0
The same with send_yield or enable_coroutine off	80.00 bytes	0
Closed by the server after a keep-alive request	80.00 bytes	0
Closed by another worker after a keep-alive request	80.00 bytes	0
WebSocket closed by the client	80.00 bytes	0
HTTP/2 closed by the client	80.00 bytes	0
Connection: close or 127.0.0.1 or SWOOLE_BASE or any other dispatch_mode	0	0

• The 6.2 branch at 522f1cc gives 80.00 bytes before and 0 after
• The 6.1 branch at bb7880e gives 0 since it caches in SWOOLE_BASE mode only

Fix

• swoole_http_server_is_ip_cache_enabled() now decides both whether the addresses are cached and whether php_swoole_http_server_onClose() is installed
• With dispatch_mode 2 the close event of a session goes to the worker that holds its entries
• send_yield is on by default and then php_swoole_server_onClose() is already installed so the change only swaps it for the HTTP variant that erases two entries and calls it
• With send_yield or enable_coroutine off nothing was installed before and a close now takes the same path as with an empty user onClose callback

Tests

• The new tests/swoole_http_server/server_addr_released_on_close.phpt fails on master with 40256 bytes for 500 connections and passes with the fix
• tests/swoole_http_server and swoole_websocket_server and swoole_http2_server and swoole_server give the same results before and after except the new test
• A few swoole_server timing tests changed between runs and failed at the same rate on both builds when they were repeated alone
• Every dispatch_mode with and without callbacks and with send_yield or enable_coroutine off gave the same replies and closed connections and onClose calls and log lines before and after
• valgrind with USE_ZEND_ALLOC=0 found no error in the master and the manager and the worker by default and with send_yield or enable_coroutine off
• 20000 client-closed connections with 4 CPUs took 2.30 s before and 2.22 s after as the median of six runs in ABBA order
• The patch applies to the 6.2 branch and gives 0 there

[NathanFreeman]

If possible, please merge this PR into the 6.2 branch. Thank You.

[NathanFreeman]

Thank you for your contribution. However, this fix needs to be merged into the 6.2 branch. I have created a new PR for this: #6292