Fix(llm): Abort orphaned requests when llm.chat() batch fails Fixes #26081

[Flink-ddd]

Purpose

Fixes #26081

This PR addresses a bug where llm.chat() pollutes the scheduler queue if a batch of requests fails validation mid-way.

The Bug:
When a batch (e.g., 4 requests) is sent to llm.chat() and one request (3) is invalid (too long), the engine enqueues the valid requests (1, 2) and then throws an exception for 3. These enqueued requests (1, 2) become "orphaned" and are not cleaned up.

The Consequence:
The next call to llm.chat() (in a recursive retry logic) picks up these orphaned requests, leading to more outputs than inputs and breaking the API contract.

The Fix:
This PR makes the batch addition in _validate_and_add_requests transactional.

1. It wraps the request-adding loop in a try...except block.
2. It tracks the request_id of each successfully enqueued request from the current batch.
3. If an exception occurs, the except block iterates through the tracked IDs and calls self.llm_engine.abort_request() on each "orphan" before re-raising the exception.

Test Plan

1. Manual Reproduction (OP's Example): Used the recursive infer() function from [Bug]: .chat() does not clean up in case of validation failure #26081 to demonstrate the bug (orphaned requests causing incorrect output counts) and verify the fix (correct output counts). See results below.
2. Automated Regression Test: Added test_chat_batch_failure_cleanup to tests/entrypoints/llm/test_chat.py. This test (a) triggers a batch failure with an invalid prompt, (b) immediately sends a new, valid batch, and (c) asserts that the output count for the new batch is correct, confirming the queue was cleaned.

Test Result

The fix is confirmed by both the manual reproduction script and the new automated test.

1. Manual Reproduction (Before vs. After)

Before Fix (on main branch): Bug Reproduced
The recursive call processed 5 real outputs instead of the expected 3, due to 2 orphaned requests.

After Fix (on this branch): Fix Verified
The recursive call now correctly processes 3 real outputs and 1 error, as the orphaned requests were successfully aborted.

2. Automated Regression Test (Pytest)

The new pytest case passes, ensuring this bug does not regress.

commands:
pytest tests/entrypoints/llm/test_chat.py -k test_chat_batch_failure_cleanup