Streaming runner with backpressure and async stage support.

Implements lazy evaluation with Task.async_stream for memory-efficient processing of large datasets.

Features

• Lazy sample generation (pull model)
• Backpressure via bounded Task mailboxes
• Async I/O-bound stages with concurrency control
• Memory bounded to O(concurrency), not O(dataset_size)

Usage

# Returns a stream (lazy evaluation)
stream = Forge.Runner.Streaming.run(pipeline, concurrency: 10)

# Consume with backpressure
samples = stream |> Enum.take(1000) |> Enum.to_list()

# Or process lazily
stream
|> Stream.each(&process_sample/1)
|> Stream.run()