*Background*

ldiskfs uses the underlying array's `stride` and `stripe_width` so that mballoc
can align allocations to full RAID rows; without them, large writes arrive
misaligned and the array services them as read-modify-write. mke2fs normally
derives both from the blkid topology, which covers a `mkfs.lustre` over a bare
md device. Two configurations fall outside what the topology can express:

• *Stacked devices* — with dm layers between the OST and the array
    (dm-linear, lvmcache), the topology is not always propagated.
• *Declustered arrays* — `optimal_io_size` describes the whole device pool
    rather than one data row, so a `stripe_width` derived from it is not the row
    mballoc should align to.

Neither is signalled today: an OST formatted without usable geometry looks
identical to one formatted with it, and the cost appears later as reduced
write throughput.

*Improvement*

`mkfs.lustre` probes the md array beneath an OST through sysfs, descending
stacked dm devices via the `slaves` links, and seeds `-E stride=/stripe_width=`
from the array's own chunk size and data-disk count. The existing
mke2fs-derived path is untouched wherever it already works, user-supplied
values always win, and if several arrays are found with disagreeing geometry
nothing is seeded.

`mount.lustre` gains a matching advisory when an OST sits on a striped array
whose filesystem carries no matching geometry, making the condition
discoverable on filesystems that already exist. Misaligned partition starts are
reported at mkfs.

With the geometry in hand, `mkfs.lustre` can also report a data row
(`data_disks * chunk`) that is not a power of two. Bulk RPC sizes are powers of
two in practice, and only a power of two divides one, so a 320 KiB row (k=5 at
64K) leaves a partial row in every RPC, at every chunk size. Unlike
stride/stripe_width this cannot be corrected once the array exists, so it is
reported while the geometry can still be chosen.

*Scope*

Userspace only, confined to `lustre/utils/libmount_utils_ldiskfs.c`. No on-disk
format, wire protocol or kernel change.

*Testing*

Built with `-Wall -Werror`; 13 mkfs and 7 mount cases on a RHEL 9.7 OSS against
live loop-backed md arrays, each cross-checked with `dumpe2fs`: raid0/4/5/6,
raid10 near and far layouts, a non-power-of-two row, user-supplied values,
dm-linear over an array, dm-linear spanning two arrays with conflicting
geometry, and aligned vs misaligned partition starts.

Declustered support is decode-verified against the array driver's layout
definitions but has not yet been run against a live declustered array.