Fixes #888.

Server-Sent Events (and other streaming responses) fail on the Fiber adapters with unable to flush. Fiber is the only adapter built on fasthttp rather than net/http, and fasthttp.RequestCtx doesn't implement http.Flusher — it can't flush the response body synchronously from within the handler, only via the deferred SetBodyStreamWriter callback. Every other adapter (chi, echo, gin, go, httprouter, mux, bunrouter, flow) exposes an http.Flusher and already works.

Approach

An internal, unexported streaming hook — no new public API:

• The sse package detects an optional bodyStreamer interface on the context, alongside the existing unwrapper/writeDeadliner capability interfaces. When present, it streams through the adapter's callback; otherwise it writes to BodyWriter directly as before.
• The Fiber v2 and v3 adapters implement it via fasthttp's SetBodyStreamWriter, passing a small writer that satisfies io.Writer, http.Flusher, and the write-deadline interface, so the shared send loop treats it uniformly.
• fasthttp stays an indirect dependency — it is not imported by the sse package or the humafiber source, so no SSE user (chi/net-http/etc.) picks up a new direct dependency.
• The SSE send loop is extracted into a shared stream() function reused by both paths.

Tests

• humafiber.TestSSE (v3) and TestSSEV2 (v2) drive SSE end-to-end through the real fasthttp stream path and assert the exact streamed output.
• The existing net/http SSE tests are unchanged and still pass, confirming no regression on the direct-write path.

Credit

Thanks to @zhouyusd for surfacing this and the first attempt in #1014, and to the reporters on #888. This PR takes a different approach that keeps fasthttp out of the core dependency graph, so it supersedes #1014.