Add AWS/Bedrock user simulator (roadmap: "User Simulator Options") + optional Bedrock agent STT/TTS & judges

[nevo-biton]

Roadmap alignment

Your roadmap lists User Simulator Options (Framework): "Introduce additional options for the user simulator, including open-source model support for greater flexibility and ease of adoption." This proposal adds a new AWS/Bedrock user-simulator option (an audio-native Amazon Nova Sonic caller), and also lets the Extended Leaderboard goal of "additional speech-to-speech systems" cover Nova Sonic. It then completes the picture so an eval can run entirely on AWS/Bedrock — extending the same AWS support to the agent-under-test (STT/TTS) and the validation judges — for users who don't have OpenAI/Deepgram/ElevenLabs/Gemini keys.

Proposed additions (implemented & live-validated; happy to split into separate PRs)

1. bedrock_s2s user simulator (the roadmap item) — audio-native Nova Sonic caller over Bedrock bidirectional streaming; the AWS analog of OpenAIRealtimeUserSimulator (reuses BotToBotAudioBridge, _build_prompt(), the end_call tool).
2. Amazon Transcribe (STT) + Amazon Polly (TTS) as assistant providers in create_stt_service/create_tts_service, wiring pipecat's AWSTranscribeSTTService/AWSPollyTTSService. No new deps (uses websockets + aiobotocore, already present via aioboto3). The agent LLM already runs on Bedrock via the LiteLLM model_list.
3. Bedrock LLM validation judges — text judges via JUDGE_MODEL; the audio judge (user_speech_fidelity) auto-skips on AWS-only setups since no Bedrock model accepts audio input.

Honest caveats

• The Nova Sonic caller needs the experimental aws-sdk-bedrock-runtime (pre-1.0, Python ≥3.12). It's isolated as an optional eva[bedrock-s2s] extra behind a lazy factory guard, so EVA's 3.11 floor and base install are unaffected.
• Speech-fidelity validation has no Bedrock equivalent (no audio-input Bedrock model), so it's skipped on AWS-only runs — not silently broken. (Tangentially related to the Metrics roadmap note that current audio judges can't assess prosody.)

Questions

Given "User Simulator Options" is on the roadmap: would you prefer this as three separate PRs (caller / agent STT+TTS / judges) — with the caller PR landing the roadmap item first — or one combined PR? A combined PR is opened for review; I'm happy to split it however you prefer.