Fix max_runs check discarding successful tool call in run_until_tool_used

[brainlid]

Problem

LLMChain.run_until_tool_used/3 with max_runs: 1 incorrectly errored with Exceeded maximum number of runs (1/1) even when the very first LLM call successfully invoked the target tool. The successful tool result was discarded in favor of the max-runs error.

The bug also affects any max_runs: N configuration where the target tool happens to be called on the Nth attempt — the success is masked by the ceiling check.

Solution

The pipeline in LangChain.Chains.LLMChain.Modes.UntilToolUsed.do_run/2 runs a series of guarded steps where each step either continues with {:continue, chain} or short-circuits with a terminal state. The original order placed check_max_runs immediately after call_llm, so the run counter was incremented and the ceiling check fired before execute_tools and check_until_tool could observe that the target tool had been called.

Reordering moves check_max_runs to run after check_until_tool. Successful termination is detected first; check_max_runs becomes a no-op on already-terminal pipeline states. A comment was added at the call site explaining the ordering constraint so it isn't innocently shuffled in the future.

Changes

• lib/chains/llm_chain/modes/until_tool_used.ex — Move check_max_runs to run after execute_tools and check_until_tool in do_run/2; add explanatory comment for the ordering constraint.
• test/chains/llm_chain_test.exs — Add regression test covering max_runs: 1 with a first-call successful tool invocation; asserts the chain returns the ToolResult rather than a max-runs error.

Testing

• Added a focused regression test in LangChain.Chains.LLMChainTest that mocks ChatOpenAI.call/3 to return the target tool call on the first attempt with max_runs: 1, and asserts both that updated_chain.last_message.role == :tool and that the returned ToolResult is non-error.
• Existing run_until_tool_used tests (including the pre-existing max-runs exceeded path) continue to cover the failure case where the target tool is never called.