Fix cached master ownership lifecycle

[ADovgalyuk]

Summary

Centralize disposal of redisMaster::cached_master in replicationDiscardCachedMaster() and transfer ownership before scheduling asynchronous client destruction.

Root cause

replicationCacheMasterUsingMyself() and replicationCacheMasterUsingMaster() called freeClient(mi->cached_master) without clearing the field. The existing safeguard in replicationCreateMasterClient() then observed the stale non-null pointer and queued the same client through freeClientAsync(), creating a double-free/use-after-free lifecycle.

Changes

• Make replicationDiscardCachedMaster() clear mi->cached_master before touching or asynchronously freeing the client.
• Route both cached-master reconstruction paths through that helper.
• Keep a defensive discard and invariant check in replicationCreateMasterClient().
• Remove direct freeClient(mi->cached_master) calls from the reconstruction paths.

Validation

• git diff --check executed while generating the commit.
• Repository CI will build KeyDB and run the configured unit-test set.

Related upstream context: Snapchat#896.