docs: correct README's description of the payment agent to match the actual implementation

[pplmaverick]

Summary

Discovered while studying the repo as a reference for building a live payment system on Arc.

The README describes the payment agent (agent.mts, run via npm run agent) as an
LLM-driven agent built with LangChain and Deep Agents that "autonomously" decides which
tools to call, with a documented fallback to a "scripted mock run" and a natural-language
query example (npm run agent -- "Buy me a quote at ...").

None of this matches the actual implementation:

• agent.mts has zero references to langchain, openai, or deepagents anywhere in
  the file (verified via full read + repo-wide grep).
• parseArgs() (agent.mts:15-32) only recognizes a single --limit <number> flag. Any
  other CLI argument — including the quoted natural-language query shown in the README
  example — is silently ignored and has no effect on execution.
• There is no OPENAI_API_KEY branch, no LLM call, and no "mock mode" anywhere in the
  script. The agent's behavior is unconditional and deterministic regardless of whether
  OPENAI_API_KEY is set.
• The actual behavior (agent.mts:91-296) is: generate an ephemeral wallet, fund it from
  BUYER_PRIVATE_KEY, deposit USDC into Circle Gateway (gateway.deposit), then run a
  fixed setInterval loop (agent.mts:259-293) that round-robins through the 4 hardcoded
  /api/premium/* endpoints at 1 request/second, auto-topping-up the Gateway deposit
  below a 0.5 USDC threshold (agent.mts:172, 199-221).
• package.json still lists @langchain/core, @langchain/openai, and deepagents as
  dependencies, but they are unused — not imported anywhere in the codebase.

This is misleading for anyone evaluating the demo's actual capabilities (e.g. assuming
it demonstrates LLM tool-use, or that the query example does something).

Changes

• Rewrote the project intro, the "Run the AI payment agent" step, and the "How It Works"
  bullet to describe the agent as what it is: a scripted, deterministic payment loop.
• Removed the non-functional "custom query" example (npm run agent -- "Buy me a quote at ..."), since arbitrary arguments are silently ignored by parseArgs().
• Removed the Prerequisites bullet about an optional OPENAI_API_KEY enabling
  "LLM-driven" behavior, since the script never reads that variable.
• No code changes — this PR only touches README.md. The unused LangChain/Deep Agents
  dependencies in package.json are a separate, follow-up cleanup (removing them, or
  actually wiring them up if LLM-driven behavior was the intent) and are out of scope
  here.

Test plan

• Diff reviewed against current agent.mts and package.json to confirm accuracy
• Markdown renders correctly (headings, code fences, bold/link syntax)