Move input coroutine management variables to TLS

[runderwo]

In c98a68e, coroutine state was moved to thread-local storage. However, the variables introduced to manage thread-local coroutine state were not themselves thread-local, which is a problem because flb_input_coro_collect() is not guarded against re-entrancy. This results in a race when multiple threads wake up from polling and run, because they then both read (e.g. in input_pre_cb_collect()) and overwrite (e.g. in flb_input_coro_collect()) the TLS key used to look up the stored coroutine state (libco_in_param_key), without synchronization.

There are almost certainly additional bugs here but this at least mitigates the race condition. The error handling logic in input_pre_cb_collect() was removed and replaced with an assert because the recovery path made no sense, and it's not clear there is in fact a sensible recovery path given the code structure.

Most likely this is related to #11488 and may be the root cause.

---

Enter [N/A] in the box, if an item is not applicable to your change.

Testing
Before we can approve your change; please submit the following in a comment:

• Example configuration file for the change
• Debug log output from testing the change

• Attached Valgrind output that shows no leaks or memory corruption was found

Fluent Bit is licensed under Apache 2.0, by submitting this pull request I understand that this code will be released under the terms of that license.

Summary by CodeRabbit

• Refactor
  • Per-thread state is now initialized once in a thread-safe, deterministic setup to improve reliability of input handling.
• Bug Fixes
  • Removed implicit lazy allocation of per-thread parameters; callers must ensure per-thread context is initialized, enabling earlier detection of uninitialized usage and more predictable teardown.