refactor: remove dead branches in async _pump_iterator and drop put_terminal

[mvallebr]

Summary

Cleanup follow-up to #33 (max_in_flight).

async _pump_iterator dead branches + dead parameters

Two isinstance(q, AsyncQueueBranch): await q.put(item) else: await q.put(item) checks for item enqueue where both arms were byte-for-byte identical — leftover from an incomplete refactor. Collapsed them into a single await q.put(item).

The materialize_before_enqueue and consumer_type parameters were also dead: every caller passed materialize_before_enqueue=False, so the True branch (which called _apply_materializer before re-iterating) was never executed. Removed both parameters and the dead branch, simplifying the pump to a single async for loop.

AsyncQueueBranch.put_terminal removal

put_terminal was functionally identical to put (same active-gated put_nowait retry loop), so it is removed and the except / finally blocks in _pump_iterator now call put for terminal markers too. AsyncQueueBranch keeps close() / active for backpressure / cancellation.

Per-test timeout

Added pytest-timeout as a dev dependency and timeout = 10 (thread method) in [tool.pytest.ini_options] so a hung streaming test fails loudly instead of stalling the whole suite. Baseline run is ~2.2s, so 10s gives ample headroom.

Scope intentionally left out — needs its own PR

SyncFanout._put_terminal (sync engine) keeps its busy-wait + get_nowait() drop on a full queue. Removing the drop without a full shutdown redesign of SyncFanout turns it into a hard hang: a terminal consumer whose output stays lazy never drains its SyncQueueIterator, the bounded queue fills up, and the pump blocks forever trying to deliver the EOF marker. The drop is the current workaround for that design gap and touching it belongs in a dedicated PR (shutdown semantics of SyncFanout, non-enqueueing abort, drain-on-cleanup).

Verification

• uv run ruff format / ruff check --select F401: clean
• uv run pytest tests/: 445 passed in 2.21s (baseline intact, timeout enabled)
• Patch coverage: 80.0% (threshold: 80%) — the one uncovered line is the on_error == OnError.STOP branch inside the async pump, which is unreachable in the current engine: STOP streams are eagerly materialized in _publish_output before the pump is ever used.