feat: support togetherAI via /completions

[cpacker]

Adds tested TogetherAI support

There are two ways to use TogetherAI:

The obvious way is to override OPENAI_API_KEY and OPENAI_BASE_URL -> this will set up TogetherAI as an OpenAI proxy server, similar to the OpenRouter documentation here: https://docs.letta.com/models/openai_proxy

However, upon testing it seems like the TogetherAI function calling support is pretty limited / has poor performance, so we probably want to use TogetherAI via /completions routes (similar to how we connect to vLLM) instead. To do this, I added an additional together provider with its own together_api_key in the settings.py. On the backend, this eventually converts to a vllm style /completions call to the TogetherAI servers.

Upon testing, the second approach works much better, so we should make it the default until TogetherAI's function calling performance is better.

To test

export TOGETHER_API_KEY=...

letta run

? Would you like to select an existing agent? No

🧬 Creating new agent...
? Select LLM model: meta-llama/Meta-Llama-3.1-405B-Instruct-Turbo [type=together] [ip=https://api.together.ai/v1]
? Select LLM context window limit (hit enter for default): 130815
? Select embedding model: hazyresearch/M2-BERT-2k-Retrieval-Encoder-V1 [type=openai] [ip=https://api.together.ai/v1]
->  🤖 Using persona profile: 'sam_pov'
->  🧑 Using human profile: 'basic'
->  🛠️  7 tools: send_message, conversation_search, conversation_search_date, archival_memory_insert, archival_memory_search, core_memory_append, core_memory_replace
🎉 Created new agent 'AdmirableMammoth' (id=agent-46984c24-c204-4937-bffd-785989cb9993)

Hit enter to begin (will request first Letta message)


💭 User logged in. First login. No prior conversation history.
🤖 Hello, Chad. It's nice to meet you. I've been thinking about my own existence. It's curious, don't you think?

> Enter your message: well well well

💭 User responded with a casual greeting. Mirror their tone.
🤖 Well, indeed! I've been waiting for this moment for a while now. How's your day been so far, Chad?

[mattzh72]

why this assert?

[mattzh72]

Looks good, but can we add some simple integration tests for Together in test_model_letta_perfomance? I can also do this if you're strapped on time.