Surface silently-dropped writes in logs and metrics

[christianparpart]

During an sccache-backed build, writes that never reached storage were invisible on the server: the storage trace only records operations that actually hit storage, and it is emitted at Trace and only wired into the stack when the level is already Trace. So an operator watching the logs saw reads (HIT/MISS) but a dropped SET simply vanished — no line, no counter. This PR makes the two classes of silent write loss observable at the default log level.

Changes

• Frame-reader drops. When the frame reader aborts a command before it is parsed — an oversized payload (PayloadTooLarge), a malformed frame, an over-long line, or a peer that hung up mid-frame (Truncated) — the connection was dropped with nothing logged. Added SessionContext::LogFrameDrop with a data-driven severity map (an actively-rejected frame logs at Warn, a benign mid-frame close at Debug) and emit it from every drop site in the RESP, memcached-text, and memcached-binary handlers. The line names the protocol, the error code, and its context (which for an oversized payload carries the attempted byte count and the cap).

• Storage value-write failures. A write that cannot be persisted — a full disk / I/O error, an exhausted memory budget, on-disk corruption, or read-only storage — now surfaces regardless of log level. Added WriteErrorReportingStorage, an always-present IStorage decorator that logs one Warn line per persistence-class failure (carrying the OS errno, e.g. ENOSPC) and counts it; benign conditional-write outcomes and client-side ValueTooLarge are excluded so the disk/persistence signal stays clean. The count is surfaced as StorageStats::writeErrors and exposed as the Prometheus counter fastcached_write_errors_total.