Add RequestSendType.ROUND_ROBIN for per-request IP round-robin

Motivation

When a host resolves to several IP addresses (e.g. a service behind DNS that returns multiple A/AAAA records or multiple backend instances), AHC today effectively pins
all traffic to a single IP:

• DNS resolution returns all IPs, but NettyChannelConnector always connects to the first address and only advances to the next one on a connection failure (failover
  only).
• The connection pool and the HTTP/2 connection registry are keyed by (scheme, host, port) — never the IP.

So with keep-alive enabled, the first reachable IP receives essentially all requests, and there is no way to spread client load across a multi-IP host's addresses. This
change adds opt-in client-side round-robin so requests are distributed evenly across all of a host's IPs.

What changed

A new config option requestSendType on DefaultAsyncHttpClientConfig:

• DEFAULT (default) — today's behavior, unchanged.
• ROUND_ROBIN — distributes requests strictly per request across all of a host's resolved IPs, while preserving TCP failover, for both HTTP/1.1 and HTTP/2.

asyncHttpClient(config().setRequestSendType(RequestSendType.ROUND_ROBIN));

How ROUND_ROBIN works

1. Resolve the host, then pick the next IP via a per-host round-robin counter and rotate the resolved address list so that IP is tried first — the remaining IPs stay in
   the list so the connector can still fail over.
2. Pin connection reuse to the chosen IP via an IP-aware partition key, so pooled HTTP/1.1 connections and multiplexed HTTP/2 connections are kept and reused per IP
   rather than per host. HTTP/2 reuse is governed by the same partition key, so the spread works there too.
3. On failover (the selected IP is down), the key is re-pinned to the IP actually connected to, so pooling/reuse stay correct.
4. maxConnectionsPerHost remains enforced per host (the permit is taken before the target IP is known).

Backward compatibility — does not change existing behavior

• DEFAULT is the default. All round-robin logic is gated behind the flag and behind a per-request override that is null in DEFAULT mode, so the default
  send/poll/connect/pool path is unchanged.
• No API break. AsyncHttpClientConfig#getRequestSendType() is added as a default method returning RequestSendType.DEFAULT, so existing external implementations of the
  interface keep compiling (source- and binary-compatible).
• Safe no-ops. Even with ROUND_ROBIN enabled, the feature does nothing (falls through to standard behavior) for: single-IP hosts, literal-IP hosts, requests with an
  explicit address, and proxied requests (the proxy host is resolved, not the target).
• Regression suite passes unchanged: connection pool, max-connections (per-host / per-thread / total), HTTP/2 (full suite), retry, and redirect tests.

Testing

• Unit: address rotation (even distribution, stable ordering, single-IP no-op, independent per-host counters), IP-aware partition key equality/re-pin, and config
  defaults / builder round-trip.
• End-to-end (loopback 127.0.0.1/2/3):
  • ROUND_ROBIN spreads connections across all three IPs; DEFAULT keeps reusing a single pooled connection — for both HTTP/1.1 and HTTP/2.
  • Failover: when the selected IP is down, the request falls over to a reachable IP and succeeds.