The fetch loop matched any error other than `Eof` (and recoverable
HTTP responses) with a catch-all `_ => break`, dropping the spawned
task and leaving the data source silently dysfunctional. Because no
one was polling the eventsource-client's stream anymore, its reconnect
logic never ran. The `None` branch also broke without notifying the
caller, so callers waiting on initialization had no signal that the
stream had closed permanently.

Replace the catch-all `break` with `continue` so the eventsource-client
gets to reconnect on the next poll, and call `init_complete(false)` in
the `None` branch so the caller can observe permanent stream closure.

Adds `streaming_source_recovers_from_non_eof_stream_error` test --
sends invalid UTF-8 to trigger `Error::InvalidLine` from the parser and
asserts the mock receives at least two requests.

Pairs with launchdarkly/rust-eventsource-client#134, which fixes the
matching state-transition gap on the eventsource-client side. Together
they restore the reconnection contract: the eventsource-client owns
reconnection, and the SDK keeps polling and trusts it.

Closes #116.

The 401 path (and any other status that `is_http_error_recoverable`
rejects) should stop the streaming fetch loop and signal init failure
to the caller. The catch-all-error change in this PR sits directly
next to that branch, so add a behavioral test exercising the 401 path
end-to-end through the fetch loop -- existing coverage was only of
the predicate function.