Behaviour for provider modules used by LlmComposer.

A provider is responsible for:

• receiving normalized LlmComposer.Message inputs,
• calling an upstream API,
• returning a normalized LlmComposer.LlmResponse.

Required callbacks

• name/0: returns the provider atom (for example, :open_ai, :google).
• run/3: executes one completion request.

run/3 receives:

• messages: user/assistant/tool messages,
• system_message: system prompt message (or nil),
• opts: provider options (model, credentials, request params, stream flag, etc.).

It must return:

• {:ok, %LlmComposer.LlmResponse{}} on success,
• {:error, reason} on failure.

Minimal implementation shape

defmodule LlmComposer.Providers.MyProvider do
  @behaviour LlmComposer.Provider

  @impl LlmComposer.Provider
  def name, do: :my_provider

  @impl LlmComposer.Provider
  def run(messages, system_message, opts) do
    # 1) validate required opts (for example, :model / auth)
    # 2) build provider request from messages + system_message
    # 3) call API and map result into {:ok, %{response: body}} | {:error, reason}
    # 4) normalize through your ProviderResponse adapter
  end
end

For consistency with built-in providers, implement a LlmComposer.ProviderResponse.* adapter so provider-specific payloads are parsed into LlmComposer.LlmResponse. If streaming is supported, also add a LlmComposer.ProviderStreamChunk.* adapter.