This makes the idedisk_release function only flush the cache on final
release; with the recent 2.6 blocklayer updates release gets called
somewhat frequently, and at times where IO is outstanding to the disk.

This bug didn't trigger before simply because ide_cacheflush_p() always
was a nop.