SDK DataLoaders 4: working around shutdown brittleness (#5337)

Introduces `RecordingStream::clone_weak`, so `DataLoader` threads
started from the SDK don't prevent the recording from flushing once
`main` goes out of scope, all the while making sure that `DataLoader`s
run to completion.


![image](https://github.com/rerun-io/rerun/assets/2910679/94bf7b16-cf9b-40ce-a190-d255328af3f8)


- Mitigates #5335 

---

Part of series of PR to expose configurable `DataLoader`s to our SDKs:
- #5327 
- #5328 
- #5330
- #5337
- #5351
- #5355

crates/re_sdk/Cargo.toml

@@ -58,6 +58,7 @@ re_types_core.workspace = true
 ahash.workspace = true
 crossbeam.workspace = true
 document-features.workspace = true
+itertools.workspace = true
 once_cell.workspace = true
 parking_lot.workspace = true
 thiserror.workspace = true
@@ -76,7 +77,6 @@ webbrowser = { workspace = true, optional = true }
 [dev-dependencies]
 re_data_store.workspace = true
 
-itertools.workspace = true
 ndarray-rand.workspace = true
 ndarray.workspace = true
 rand.workspace = true