fix(agents): stop agents at max_tool_calls / max_iterations - #2365

`per_user_react_agent_workflow` built `ReActAgentGraph` without `raise_on_parsing_failure` and `use_native_tool_calling`, though `PerUserReActAgentWorkflowConfig` inherits both from `ReActAgentWorkflowConfig`. Setting either had no effect: native tool calls were never bound to the LLM, and a parsing failure always raised. Both are now passed from the config, as `react_agent` does.

Tests in `tests/agent/test_react.py`, each run for `react_agent` and `per_user_react_agent`:
- `raise_on_parsing_failure: false` returns the parsing error as the answer
- `use_native_tool_calling: true` runs the model's structured tool call and returns the final answer

The `per_user_react_agent` cases fail on develop; the `react_agent` cases pass and keep both paths in step. `nvidia_nat_langchain/tests`: 766 passed, 13 skipped. yapf and ruff pass.

This touches the same `ReActAgentGraph(...)` call as #2365 (which adds `max_tool_calls`) on different lines; the two merge cleanly.

Closes #2366

## By Submitting this PR I confirm:
- I am familiar with the [Contributing Guidelines](https://github.com/NVIDIA/NeMo-Agent-Toolkit/blob/develop/docs/source/resources/contributing/index.md).
- We require that all contributors "sign-off" on their commits. This certifies that the contribution is your original work, or you have rights to submit it under the same license, or a compatible license.
  - Any contribution which contains commits that are not Signed-Off will not be accepted.
- When the PR is ready for review, new or existing tests cover these changes.
- When the PR is ready for review, the documentation is up to date with these changes.




## Summary by CodeRabbit

* **New Features**
  * Per-user ReAct workflows now honor configuration for handling unparseable model output and using native tool calling. When enabled, structured tool calls can invoke configured tools, while parsing behavior follows the selected setting. This brings per-user workflows in line with the corresponding options available in standard ReAct workflows.

Authors:
  - Somay kousis (https://github.com/Somay-kousis)

Approvers:
  - Will Killian (https://github.com/willkill07)

URL: #2367

When no `Final Answer:` marker was streamed, ReAct's `stream_fn` fell back to the text it had buffered across every agent turn. After a tool call, a direct answer was streamed behind the earlier `Thought:`, `Action:` and `Action Input:` text, and with `raise_on_parsing_failure: false` a parsing failure streamed the raw reply instead of the parsing error. `single_fn` returns the agent node's `final_answer` in both cases.

`stream_fn` now streams with `stream_mode=["messages", "updates"]`, as the tool-calling agent's `stream_fn` already does, keeps the `final_answer` from the agent node's update, and yields it in that fallback. Token streaming after a `Final Answer:` marker is unchanged, and the buffered text is still the fallback when the graph sets no answer.

Behavior change for streaming clients: in those two cases they now get the same text as `single_fn`, including the parsing error that `react-agent.md:96` describes, instead of the raw reply.

Tests in `tests/agent/test_react.py`:
- A parametrized test runs the same scripted replies through `single_fn` and `stream_fn` and requires the same answer: a direct answer (also wrapped in `<think>`), a tool call then a direct answer, and a parsing failure on the first turn and after a tool call. The last three fail on develop; the first two pass there and guard the fallback.
- The `_stream_fn` tests that mock `graph.astream` now yield `("messages", (chunk, metadata))` pairs, the shape `astream` yields for a list of stream modes. Only the mocks' event shape changed; their assertions are unchanged.
- `_run_react_agent` (from #2365) takes `raise_on_parsing_failure`, with the default unchanged.

`nvidia_nat_langchain` tests: 827 passed, 13 skipped on Python 3.11, 3.12 and 3.13. yapf and ruff pass.

Follow-up to #2380, which left this out of scope.

Closes #2432

## By Submitting this PR I confirm:
- I am familiar with the [Contributing Guidelines](https://github.com/NVIDIA/NeMo-Agent-Toolkit/blob/develop/docs/source/resources/contributing/index.md).
- We require that all contributors "sign-off" on their commits. This certifies that the contribution is your original work, or you have rights to submit it under the same license, or a compatible license.
  - Any contribution which contains commits that are not Signed-Off will not be accepted.
- When the PR is ready for review, new or existing tests cover these changes.
- When the PR is ready for review, the documentation is up to date with these changes.






## Summary by CodeRabbit

* **Bug Fixes**
  * Improved streamed agent responses when they don’t include a “Final Answer” marker. The completed answer is now returned when available, with buffered text used as a fallback, helping streamed results match single-response answers across direct replies, tool calls, and parsing failures.

Authors:
  - Hesam Shams (https://github.com/hesam-shams)

Approvers:
  - Will Killian (https://github.com/willkill07)

URL: #2433