Loosen connect timeouts for slow links

Bump the default per-host TCP connect timeout 2s -> 4s (satellite links saw no node respond under 2s on push), and give the cold replica-discovery probe its own more patient 10s budget — that first "which nodes serve this repo" contact pays a cold DNS+TLS handshake to a possibly-distant entry LB and has no failover to fall back on, so a tight budget fails the whole clone/fetch. Replica failover keeps the short default so dead nodes are still skipped quickly.

An explicit ENTIRE_CONNECT_TIMEOUT_SECONDS now overrides every connect timeout (discovery included) via os.LookupEnv, instead of being floored.

git-remote-entire's auth client (cluster /.well-known discovery + token exchange) also uses the patient discovery budget: it talks to single control-plane hosts with no failover.