[Enhancement] Add pipeline composition helpers for chaining STT, Intelligence, and TTS operations

[deepgram-robot]

Summary

Add convenience methods for composing multi-step Deepgram API pipelines — STT → Audio Intelligence → TTS — as a single fluent chain, reducing boilerplate and making common multi-API workflows expressible in a few lines.

Problem it solves

Developers frequently need to chain Deepgram APIs together: transcribe audio, run intelligence analysis (sentiment, topics, entities), then synthesize a summary or response via TTS. Today this requires manually wiring together three separate API calls, handling intermediate data transformation, and managing error propagation across the chain. A pipeline composition helper would make these common patterns as simple as:

result = await (
    deepgram.pipeline()
    .transcribe(audio_source, model="nova-3")
    .analyze(sentiment=True, topics=True, summarize=True)
    .speak(model="aura-2", voice="asteria")
    .run()
)
# result.transcript, result.intelligence, result.audio

This mirrors the operator-overloaded pipeline composition pattern becoming standard in AI frameworks, making Deepgram's multi-API workflows competitive in ergonomics.

Proposed API

from deepgram import DeepgramClient, Pipeline

client = DeepgramClient()

# Fluent pipeline builder
pipeline = (
    client.pipeline()
    .transcribe(source={"url": "https://example.com/audio.wav"}, model="nova-3", smart_format=True)
    .analyze(sentiment=True, topics=True, summarize="v2")
    .speak(text_from="summary", model="aura-2-en", encoding="mp3")
)

result = await pipeline.run()
# result.transcript: PrerecordedResponse
# result.intelligence: AnalyzeResponse  
# result.audio: bytes (TTS output)

# Or partial pipelines
stt_result = await client.pipeline().transcribe(source=audio).run()
stt_intel = await client.pipeline().transcribe(source=audio).analyze(sentiment=True).run()

Acceptance criteria

• Pipeline class supports .transcribe(), .analyze(), and .speak() steps
• Steps can be composed in any valid order (transcribe → analyze, transcribe → speak, transcribe → analyze → speak)
• Error in any step propagates with clear context about which step failed
• Intermediate results are accessible on the pipeline result object
• Documented with usage example
• Compatible with existing API (does not change any existing client methods)

---

Raised by the DX intelligence system.