Features/streaming

[mccrindlebrian]

This PR provides functionality for validator and miner streaming.

Key Points

1. Removal of docs/stream_tutorial to clean up repo
2. Deprecation of BaseMinerNeuron, into a new BaseStreamMinerNeuron that changes the requirements for miner forward method
3. Validator forward method now includes a handle_response function, which handles the async calls to the appropriate miners and acquires their completed StreamPromptingSynapses'
4. Langchain miners (HF) now store the response in a background Queue handled by a CustomTextStreamerIterator. Langchain miners are thrown onto a new thread where generations are done in batches of config.neuron.streaming_batch_size. When the generation is done, the threads are closed, and the Queue is dumped. In cases where timeout is reached, the Queue will not be empty, and on subsequent calls to the miner they will reply with tokens generated from the previous uncompleted request. This is undesirable, so we forcibly clear this Queue.
5. load_pipeline can now return streamer classes like the one mentioned in point 4
6. The prompting.protocol holds the StreamPromptingSynapse and is used to decode streamed chunks.
7. We include a new script called client.py to ping a miner on testnet without running a validator.
8. A new MockDendrite and MockStreamMiner are included in the mock.py file for testing purposes.

Block Diagram

This is a rough idea of what is happening during streaming for the most complicated case (langchain)

Limitations

1. get_openai_callback from langchain doesn't seem to integrate easily with the async calls that occur in the _forward method. This callback is used to track overall costs and openai information. I was unable to get this to work within a satisfactory amount of time, so therefore openai based miners do not use the callback anymore. This is a direction for future work.

[steffencruz]

Still needs a README