Fix redundant direct I/O tail rewrite in WritableFileWriter

[taimoor789]

Summary

Under O_DIRECT, WriteDirect()/WriteDirectWithChecksum() pad an unaligned tail with zeros, write it, and park the tail bytes at the front of buf_ via RefitTail() so a future Append() can glue more data onto them before the next write. Neither function marked that parked tail as already written, so a later Flush(), Sync(), or Close() call with no intervening Append() would re-issue a PositionedAppend() for the same already-written bytes -- not a correctness issue, since the same bytes just get rewritten at the same offset, but wasted I/O.

This adds a direct_io_buf_is_clean_tail_ flag: set after RefitTail() when the buffer holds only an already-written tail, cleared the moment Append() adds new data. Flush()'s direct-I/O path now also checks this flag and skips the redundant write.

Fixes #12168, following the fix direction @ajkr suggested on the issue: track whether the buffer is dirty or clean and skip the write when it's clean.

Test Plan

Added WritableFileWriterTest.DirectIOSkipsRedundantTailRewrite (util/file_reader_writer_test.cc), which counts PositionedAppend() calls across Flush()/Sync()/Close() sequences with no intervening Append() and asserts no redundant write occurs, plus that appending more data still glues onto the tail correctly and final file content is byte-correct. Fails on unfixed code, passes with the fix. Full file_reader_writer_test suite (114 tests) and make check pass; make check-sources clean.