Fast Tier OSTs and PCC-RO caches hold copies of data that can be dropped when
space runs low. Nothing reclaims them by policy today.

Add lfu-purge, as described in the lfu-purge HLD by Andreas:

• One process per cache keeps free space and free inodes between free_min and free_max.
• A scanner thread finds candidates with LFU scans: llapi_scan_device() on a Fast Tier OST, llapi_scan_namespace() on a PCC-RO dataset. It keeps them in a weighted-age histogram, so the main thread can purge from the old end without waiting for a scan.
• The scan rate follows free space.
• Policy classes use lfs find predicates, with weights and used_min/used_max.
• Before a purge, each candidate is checked again: its age, pins, opens, and for a fast mirror that another mirror is in sync.
• For a striped component, the OST with the first stripe owns the purge.

Also add llapi_mirror_split(): the daemon must drop a mirror as lfs mirror split -d does, and that code is static in lfs.c now.

Phase 1 of the HLD: on an OSS, the daemon needs a client mount. Setting up PCC-RO from mount -o pcc waits for LU-21082.