perf(extract): borrow instead of clone + parallelize channel browse

Four allocation/latency wins in the hot extraction paths. Behavior
preserved — all 120 in-crate tests green, clippy clean.

- stream_extractor: borrow streamingData out of the owned android + ios
  player responses instead of deep-cloning the largest subtree of each
  response. A shared Value::Null static backs the absent case.
- stream_extractor: merge_formats returns borrowed &Value format objects
  rather than cloning each one (~20-40 per video) — they are only read
  (.get()) downstream.
- channel: fetch the Home and Videos tabs concurrently via thread::scope
  so a channel open costs one round-trip of latency, not two. Home stays
  mandatory; the Videos tab stays best-effort (now also resilient to a
  panicked worker thread).
- downloader: build the response body with String::from_utf8 (in-place on
  the valid-UTF-8 common case) instead of from_utf8_lossy, which always
  copies; the lossy fallback for invalid bytes is preserved.

src/downloader/default_impl.rs

@@ -117,7 +117,14 @@ impl Downloader for ReqwestDownloader {
                 "response body exceeded cap {MAX_BODY_BYTES}"
             )));
         }
-        let body = String::from_utf8_lossy(&buf).into_owned();
+        // Reuse the body buffer on the valid-UTF-8 path (the overwhelming
+        // common case): String::from_utf8 reinterprets the Vec in place,
+        // whereas from_utf8_lossy always allocates + copies. Fall back to
+        // lossy only on genuinely invalid bytes, preserving U+FFFD behavior.
+        let body = match String::from_utf8(buf) {
+            Ok(s) => s,
+            Err(e) => String::from_utf8_lossy(e.as_bytes()).into_owned(),
+        };
 
         Ok(Response::new(code, message, headers, body, url_after_redirects))
     }