Optimize bucket listing for 10K-100K objects

- Shallow listing: read per-directory _index.json once for eTags instead
  of N serial .meta.json reads. Validate prefix for path traversal and
  verify normalized target stays within bucket root.
- Recursive listing: cache full per-directory index during the walk so
  each _index.json is parsed at most once per call.
- Per-bucket listing cache with 5s TTL and per-bucket rebuild mutex.
  Invalidated on put/delete/copy/metadata/tags/multipart-complete.
  Pagination uses partition_point for O(log n) start lookup.
- UI stream endpoint now actually streams via mpsc + Body::from_stream
  instead of buffering into a Vec<String>. Cancels producer on client
  disconnect.
- UI JSON endpoint honors delimiter=/ and returns common_prefixes.
- run_blocking wrapper dispatches sync filesystem work via
  block_in_place on multi-threaded runtimes, falls back to inline on
  current-thread runtimes (unit tests).

Cargo.lock

@@ -2740,6 +2740,7 @@ dependencies = [
  "tempfile",
  "tera",
  "tokio",
+ "tokio-stream",
  "tokio-util",
  "tower",
  "tower-http",
@@ -4193,6 +4194,17 @@ dependencies = [
  "tokio",
 ]
 
+[[package]]
+name = "tokio-stream"
+version = "0.1.18"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "32da49809aab5c3bc678af03902d4ccddea2a87d028d86392a4b1560c6906c70"
+dependencies = [
+ "futures-core",
+ "pin-project-lite",
+ "tokio",
+]
+
 [[package]]
 name = "tokio-util"
 version = "0.7.18"