feat: add @fast.maker decorator for reliable multi-step workflows

[lucidprogrammer]

Description

Adds @fast.maker decorator implementing MAKER reliability patterns from "Solving a Million-Step LLM Task with Zero Errors".

Enables zero-error multi-step workflows through k-threshold voting and red-flag validation.

Key Features

• First-to-ahead-by-k voting: Configurable margin requirement for consensus
• Red-flag filtering: Discard suspicious responses (too long, malformed) before voting
• Match strategies: exact, normalized
• Transparent results: Access vote counts, margins, and convergence status via last_result

Usage

@fast.agent(name="worker", model="claude-3-haiku-20240307")
@fast.maker(
    name="reliable_worker",
    worker="worker",
    k=3,                      # Require 3-vote margin
    match_strategy="normalized",
    red_flag_max_length=100,  # Discard verbose responses
)
async def main():
    async with fast.run() as agent:
        result = await agent.reliable_worker.send("Classify this")

When to Use

• ETL pipelines with thousands of transformations
• Code migration across many files
• Document processing at scale
• Any task where errors compound over many steps

Testing

8 integration tests covering voting, red-flagging, and match strategies
Example in examples/workflows/maker.py

Checklist

• Code follows project style guidelines
• Tests added and passing
• Example implementation included
• No breaking changes

[evalstate]

This is a very very cool feature. Couple of quick questions - I can imagine using this quite a lot - but not read the paper yet...!

1. I often use what I call "responders" - agents with a set of template messages that don't retain history - is it part of the design of this to use the same context, or would using responders help?
2. Is there any overlap with using parallel as part of this too?
   I've got PR on the docs I'd appreciate if you could review. TY!

[iqdoctor]

@lucidprogrammer https://discord.gg/wugKkpaR8e