feat(sse): reconnecting client with Last-Event-ID replay and backoff

[OmarAlJarrah]

Adds a reconnecting Server-Sent Events client to dexpace-sdk-core, layered on the existing sans-io SSE parser. The parser already turns a byte stream into events; this adds the connection lifecycle that long-lived SSE consumers need — transparent reconnection with Last-Event-ID replay and backoff — with no new runtime dependency.

What's added

• SseConnection and AsyncSseConnection (http.sse) iterate events across reconnections. Each (re)connection is dispatched through a Pipeline / AsyncPipeline (or a bare send-callable), so retry, auth, redirect, and tracing apply per connection.
• A public retry accessor on SseParser / AsyncSseStream, so a reconnecting client can read the server's sticky reconnect hint even from a retry:-only frame that emits no event.
• The pipeline-dispatch protocols (SyncPipelineLike / AsyncPipelineLike and the send-callable aliases) are lifted from the paginator into pipeline.dispatch so the paginator and the SSE client share one definition. The pagination package keeps re-exporting them, so its public surface is unchanged.

Behavior

• Reconnects on both a transient transport error and a clean end-of-stream (browser EventSource semantics); a non-success HTTP status is a permanent failure and raises HttpResponseError without reconnecting.
• Replays the last seen event id via the Last-Event-ID header; an explicit empty id clears it.
• Backoff uses the server retry: value as its base (falling back to a configurable default), grows exponentially across consecutive failures, is jittered upward only (so a reconnect never wakes before the suggested interval), and is capped. The failure counter resets once a connection yields an event.
• max_reconnects bounds consecutive failures; on exhaustion it raises, chaining the last transport error as the cause.
• Cancellation propagates while the response is released through the shielded-cleanup convention, so a cancelled consumer never leaks the transport handle.
• The caller ends the stream by breaking the loop (for example on a data: [DONE] sentinel) or via the (async) context manager.

Verification

ruff check, ruff format --check, mypy --strict (216 source files), and the full pytest suite (1013 passing) are green. New tests cover reconnect-on-drop and clean-EOF, Last-Event-ID replay and empty-id clearing, server-retry: backoff, upward-jitter direction, exponential growth, failure-counter reset, max_reconnects exhaustion with cause chaining, non-success raises, and async cancellation with deterministic cleanup — for both the sync and async clients.