Fix AdaptiveBatcher timeout calculation (#2093)

Before the fix, the timeout calculation subtracted the passed time from the
timeout calculated for the previous round instead of subtracting it from the
configured timeout value, causing time to pass quicker than expected.

Author: hanlaur

mlserver/batching/adaptive.py

@@ -136,7 +136,7 @@ async def _batch_requests(self) -> AsyncIterator[BatchedRequests]:
 
                     # Update remaining timeout
                     current = time.time()
-                    timeout = timeout - (current - start)
+                    timeout = self._max_batch_time - (current - start)
             except asyncio.TimeoutError:
                 # NOTE: Hit timeout, continue
                 pass