[Bugfix] Propagate shared cache load failures

[dante159753]

Purpose

Prevent shared Cache buffers from being reused while an asynchronous backend still accesses them, and propagate backend load failures to every rank waiting on the same shared entry.

Modifications

• replace the shared cache ready bit with LOADING/READY/FAILED state and retain the backend failure status
• keep a cache buffer reference in backend task shards through Posix AIO/psync and Mooncake task lifetimes
• make non-owner loads consume the shared success or failure result with a bounded wait
• drain submitted backend tasks before releasing cache task state
• add unit coverage for cross-mapping failure propagation, retry after references are released, backend lifetime retention, timeout, and cleanup paths

User impact

When one rank owns a shared Cache load and the backend fails or times out, peer ranks now observe the same failure instead of waiting indefinitely or treating incomplete KV data as ready. Backend IO also cannot outlive and silently overwrite a reused cache buffer.

Verification

• git diff --check
• clang-format 20 check on changed lines
• codespell 2.4.1 on changed files
• A2, latest origin/develop: Debug/Ascend build passed; ucmstore.test 89/89 passed
• A2 ucmshared.test: 35/37 in the full run; the two logger ordering tests passed 2/2 when rerun together in isolation
• A2 DeepSeek-V2-Lite, TP2, Cache|Posix, AIO, O_DIRECT, layerwise, recompute: after truncating one Posix KV file and injecting delayed reads, only the owning rank issued backend loads, the peer consumed the shared failure, recovery completed, and generated token IDs matched the baseline