storage.Writer took an assumption that CloseWithError() could be called more than once, and was thread-safe with respect to concurrent Write(), Flush(), and Close() calls. This was not honored in the refactor in #12422.

Modify Writer so that it is thread-safe to provide these behaviors, and support repeated Close() and CloseWithError() calls.

To address this, we start the sender goroutine earlier, and gather the first buffer in that goroutine. It's possible that some workloads which gather less than one buffer worth of data with a sequence of small writes will observe a performance hit here, since those writes used to be direct copies but will now be a channel ping-pong. If that's an issue, it could be improved by wrapping the buffer in a mutex and doing more explicit concurrency control. Throughput for large writes is
not affected.

Fixes #12752