gazette: surface fragment store diagnostic on FRAGMENT_STORE_UNHEALTHY

[jgraettinger]

Problem

The broker's AppendResponse and FragmentStoreHealthResponse both carry a store_health_error string — the actual diagnostic from the failing fragment store — populated when the status is FRAGMENT_STORE_UNHEALTHY. But the gazette crate collapsed every non-OK status into Error::BrokerStatus(status), which holds only the enum value. The response (and its store_health_error) was discarded, so upstack callers saw only the generic "unexpected broker status: FragmentStoreUnhealthy".

Fix

• Add a dedicated Error::FragmentStoreUnhealthy(String) variant (Display: "fragment store is unhealthy: {0}"). A new variant, rather than a field on BrokerStatus, avoids churning the ~15 call sites that pattern-match BrokerStatus(broker::Status::X) as a tuple.
• Classify it as non-transient in is_transient(), matching prior behavior (the status means the store has failed its health check for an extended period).
• The two response types carrying store_health_error (append and fragment_store_health) now special-case FRAGMENT_STORE_UNHEALTHY and move that string into the new error before the response is dropped.

Bonus

This also fixes the storage_mappings.rs GraphQL health-check consumer. Its existing Ok(Err(err)) => Some(err.to_string()) arm previously produced the useless generic status string; it now surfaces the store's real diagnostic.

Retry semantics are unchanged: the append Stream still retries on its own loop regardless of is_transient().

cargo check passes for gazette and dependent crates (publisher, shuffle, dekaf, control-plane-api, migrate); cargo fmt applied.