Download inputs off the event consumer loop

[philandstuff]

Currently, we download inputs (via URLPath's convert() method) in
_prepare_payload, in the main Worker event consumer loop. This is bad, because
downloading inputs is a blocking operation and we cannot process other events
while downloads are in progress.

This is a substantial restructuring:

• the event consumer loop now only processes events from the child; prediction
  requests and cancel requests from the http server are now processed directly.
• there are new concurrent.future.ThreadPoolExecutor instances for preparing
  predictions and downloading inputs.
  • as part of this, this commit also supports downloading inputs concurrently
    rather than serially. I have hardcoded a maximum of 8 concurrent downloads
    based on the ThreadPoolExecutor size. I picked this number arbitrarily.
• when the web server receives a request to start a prediction, we submit a task
  on the prediction start executor, which in turn submits downloads on the input
  download executor. Once all inputs are downloaded, we send the prediction
  request to the child.
• cancel requests are handled similarly, but as there is no input download this
  is much simpler.
• all calls to self._events.send() now happen within the http server async event
  loop, not the Worker ThreadPoolExecutor. (This is maybe not strictly correct,
  as we're calling blocking i/o from an async event loop, but it's a local
  Connection object and it's a tiny amount of data)