fix(runner): reuse connections to runtimes across executions

[HarshMN2345]

The executor opened a new connection to the runtime for every execution (curl_init() per call). Under load the closed connections pile up in TIME_WAIT in the executor's network namespace, connect() gets expensive and eventually fails with EADDRNOTAVAIL.

This keeps idle curl handles per runtime hostname in each worker and reuses them (curl_reset() keeps the open connection), for v2 and v5 executions. A handle is only returned to the pool after a successful request, at most 128 are kept per runtime, and handles of runtimes not used for 60 s are closed. Runtime hostnames are random per runtime, so a recreated runtime never gets an old connection, and libcurl replaces connections that the runtime closed (tested with a runtime restart).

A shared connection cache (curl_share_init_persistent with CURL_LOCK_DATA_CONNECT) doesn't work here: Swoole's curl hook runs each curl_exec() in its own multi handle, which keeps at most 4 connections, and moving sockets between multi handles crashed Swoole under load.

Refs appwrite/appwrite#14094

Verification: real createExecution() against a keep-alive runtime stand-in, 20000 executions with 100 coroutines, 4 runs back to back:

	before	after
exec/s per run	3245, 332, 355, 1387	~28000 each
new connections per run	20000	100
TIME_WAIT per run	+20000	0
errors	300 per run in runs 2-3 (curl error 7)	none

Unit tests, phpstan, pint and rector pass.