Raise on model responses without text - #534

…ogle#534

Rebasing onto main surfaced a behavioural conflict that merged cleanly as
text. google#534 added an `except InferenceRuntimeError: raise` pass-through to the
parallel result loop, which returns the provider's own diagnostic unwrapped.
That pass-through runs before the draining logic here, so once
_process_single_prompt started wrapping failures as InferenceRuntimeError,
every parallel failure short-circuited and the completed chunks were discarded
again, which is the exact behaviour this branch set out to fix.

Failures are now collected uniformly and the pool is allowed to drain. Which
error is raised then depends on whether there is anything to preserve:

- Nothing completed: re-raise the provider's own InferenceRuntimeError
  unchanged, so a blocked prompt or an exhausted token budget still reports its
  own reason rather than being buried a level down, with the drained
  bookkeeping attached to it.
- Some chunks completed: raise the aggregate error carrying partial_results and
  failed_indices, with the first underlying error kept as `original` and quoted
  in the message.

Both test suites pass together: the no-text tests added by google#534 and the
partial-result tests on this branch.

…ogle#534

Rebasing onto main surfaced a behavioural conflict that merged cleanly as
text. google#534 added an `except InferenceRuntimeError: raise` pass-through to the
parallel result loop, which returns the provider's own diagnostic unwrapped.
That pass-through runs before the draining logic here, so once
_process_single_prompt started wrapping failures as InferenceRuntimeError,
every parallel failure short-circuited and the completed chunks were discarded
again, which is the exact behaviour this branch set out to fix.

Failures are now collected uniformly and the pool is allowed to drain. Which
error is raised then depends on whether there is anything to preserve:

- Nothing completed: re-raise the provider's own InferenceRuntimeError
  unchanged, so a blocked prompt or an exhausted token budget still reports its
  own reason rather than being buried a level down, with the drained
  bookkeeping attached to it.
- Some chunks completed: raise the aggregate error carrying partial_results and
  failed_indices, with the first underlying error kept as `original` and quoted
  in the message.

Both test suites pass together: the no-text tests added by google#534 and the
partial-result tests on this branch.