Add more checks for buffer corruption on startup

### Is your feature request related to a problem? Please describe.

Currently when starting up Fluentd outputs, we try to check if each buffer chunk
is non-empty, and if it has some bytes, we assume it contains valid data.

It turned out that this operation model has a few issues:

* If the data was indeed corrupted, the behavior is undefined. It likely causes
many kinds of errors in various parts of the pipeline.
* It is also hard to tell which chunks was corrupted from td-agent.log.
This is important because users probably want to recover the lost data.

We should perform more rigorous buffer checks on startup,
so that Fluentd can handle corrupted chunks gracefully.

### Describe the solution you'd like

* Perform more sanity checks on buffer chunks on startup.
* Emit more error logs regarding the corrupted chunks.

### Describe alternatives you've considered

N/A

### Additional context

_No response_