refactor: use curl_global_init_mem with pg allocators

[steve-chavez]

Uses curl_global_init_mem(CURL_GLOBAL_DEFAULT, net_malloc, net_free, net_realloc, net_strdup, net_calloc) to keep better track of memory allocation by curl.

The aim is to fix #216.

Unfortunately this currently segfaults when reaching 2 iterations:

-- do twice
select net.http_get('http://localhost:8080/pathological?status=200') from generate_series(1,10);
select net.http_get('http://localhost:8080/pathological?status=200') from generate_series(1,10);

-- or once (this will do two iterations since the batch size defaults to 200)
select net.http_get('http://localhost:8080/pathological?status=200') from generate_series(1,300);

The logs:

2025-08-14 17:55:08.571 -05 [643019] DEBUG:  net_realloc(palloc): a((nil)), sz(32)
2025-08-14 17:55:08.571 -05 [643019] DEBUG:  net_malloc: sz(32)
2025-08-14 17:55:08.665 -05 [643013] LOG:  background worker "pg_net 0.19.5 worker" (PID 643019) was terminated by signal 11: Segmentation fault

Using gdb (via convenience sudo xpg gdb) will show that this will always fail when hitting curl functions:

Curl_attach_connection
Curl_llist_count
Curl_conn_is_alive

Which are called when using curl_multi_socket_action in our codebase in

pg_net/src/worker.c

Lines 320 to 336 in a7792bf

	for (int i = 0; i < nfds; i++) {
	if (is_timer(events[i])) {
	EREPORT_MULTI(
	curl_multi_socket_action(worker_state->curl_mhandle, CURL_SOCKET_TIMEOUT, 0, &running_handles)
	);
	} else {
	int curl_event = get_curl_event(events[i]);
	int sockfd = get_socket_fd(events[i]);
	EREPORT_MULTI(
	curl_multi_socket_action(
	worker_state->curl_mhandle,
	sockfd,
	curl_event,
	&running_handles)
	);
	}

Unfortunately this is quite hard to debug, it might be a bug in curl itself.

[steve-chavez]

Unfortunately this is quite hard to debug, it might be a bug in curl itself.

I think this is because some curl allocations are done in sub memory contexts and the curl_global_init_mem is now causing some frees and mallocs to happen at the top level.