Improve indexing performance: batch stat()s with io_uring

[dcolascione]

Add support for using io_uring to batch stat() calls when scanning maildir directories. This approach reduces the number of individual syscalls by processing stats in batches of up to 16384 files at a time; it also allows the kernel to continue doing kernel-internal stat()s while we do indexing work.

Performance impact is moderate but noticeable on a real-world maildir with ~490k messages:

Without io_uring: 17.5s real time (0.75s user, 5.6s sys) With io_uring: 15.4s real time (0.74s user, 8.4s sys)

The higher sys time with io_uring reflects the batch processing happening in kernel space rather than repeated userspace->kernel transitions.

Only enable io_uring for maildir directories since:

1. They're the only directories likely to be large enough to benefit
2. This ensures the io_uring instance isn't used concurrently

The feature can be disabled at runtime via MU_DISABLE_IO_URING=1 or at build time via -Diouring=disabled. Requires liburing >= 2.3.

[djcb]

Looks interesting. Will take a bit longer to review this (I'm not too familiar with io_uring), I don't think before 1.12.9 (which should be in the coming week)

The performance improvements are modest, so we'll have to weigh that against the added complexity. Anyway, thanks for doing this!