[thomasjm]

The FUSE writeback cache makes the kernel keep its own file sizes, so a file grown from outside the mount reads back truncated while its inode is cached, and no FUSE notify can fix that. This splits writeback into two settings:

• --writeback-cache / Tuning::writeback_cache: the kernel cache, now off by default.
• --write-pipeline / Tuning::write_pipeline: send Twrites in the background, on by default.

Without the cache the pipeline has to keep reads, stats and write order correct itself: it waits for in-flight writes before getattr/read/lookup, keeps overlapping and O_APPEND writes in order, and merges small sequential writes while all slots are busy (4k writes at 2 ms RTT: ~8 MB/s without merging, ~110 MB/s with).

Also fixes two flaky failures in the diod tests: a hang when an unmount found the mount briefly busy, and mounts failing when another test took diod's port first.

Breaking: Tuning::writeback is replaced by the two fields above. --writeback still works but is deprecated. New: mount_with_invalidations takes the invalidation request and reply streams (used by the tests).