adds more robust arg parsing

[MrDiggles2]

Noticed that with some "weaker" models like openai-gpt-oss-20b tend to over-escape JSON for tool call arguments, e.g. {"query": \\"foo\\"} instead of {"query": "foo"}

This PR adds a new function to repair the JSON if the initial parse fails.

Also ran make format_diff

[bnarasimha21]

Thanks for this — the intent is right and the repair does fix the openai-gpt-oss-20b over-escape case. But the blanket replace('\\"', '"') introduces a regression worse than the bug it fixes:

1. Truncated payloads now execute with silently corrupted arguments. A payload cut off right after a legitimate in-string escape — e.g. {"note": "5\"} (realistic max_tokens truncation) — fails the first parse, then "repairs" to valid JSON {"note": "5"} with the quote deleted. Before this PR that payload was skipped with a warning; now the tool runs with wrong args. Worse example: {"a": "\"} repairs to {"a": ""} — the entire value erased. Suggest guarding the repair so it only triggers when the payload looks wholly over-escaped (e.g. starts with {\"), and adding {"note": "5\"} as a "must skip, not mis-parse" regression test.

2. Streaming has the same bug with a quieter failure. _stream puts raw argument strings into tool_call_chunks, and langchain-core's chunk assembly (parse_partial_json) silently degrades over-escaped args to args: {} — the tool call lands in tool_calls, not invalid_tool_calls, and executes with empty arguments (verified on langchain-core 1.4.8, the locked version). If the repair is worth doing in _generate, the streaming path needs it more.

Also a heads-up: most of this diff is make format_diff churn, which will conflict with #32.