LU-18556 hsm: optimize llog record modification

LU-18556 hsm: optimize llog record modification

This commit introduces a new llog modification mechanism for HSM
operations to address inefficiencies caused by prior reliance on
catalog processing. The new approach directly modifies llog record,
eliminating the need for catalog-based processing and reducing
latency.

Key changes include:
* Replacing the hsm_action_item (HAI) with a full in-memory llog
 record representation, increasing memory usage by ~80 bytes per
 record but removing the need for a dedicated llog cookie hash
 table.
* Unifying the coordinator's read/store logic for HAI data into a
 single in-memory item shared by mdt_hsm_agent_send() and
 mdt_hsm_add_hsr(). This reduces memory allocation steps: only one
 cdt_agent_req allocation is now required during llog read
 operations, eliminating subsequent allocations/copies.

Performance results on VMs 2 MDTs/2 OSTs/2 Clients no-op copytool:
Test 1 (1M archive requests): 572s -> 187s (~3 times faster)
Test 2 (1M archive + 1M queued): 558s -> 392s (~1.4 times faster)

HPE-bug-id: LUS-12583
Signed-off-by: Alexander Boyko <alexander.boyko@hpe.com>
Change-Id: I4b6e697bc3b1f0cf2c76f5433b49affbc933c653
Reviewed-on: https://review.whamcloud.com/c/fs/lustre-release/+/57428
Tested-by: jenkins <devops@whamcloud.com>
Tested-by: Maloo <maloo@whamcloud.com>
Reviewed-by: Nikitas Angelinas <nikitas.angelinas@hpe.com>
Reviewed-by: Etienne AUJAMES <eaujames@ddn.com>
Reviewed-by: Oleg Drokin <green@whamcloud.com>