[kean]

Four small caching fixes from the DefectRepros batch:

• Concurrent writes to different ImageCache limits could overwrite each other: each setter went through Cache.conf's separate get and set, so a countLimit write on one thread could put back the old costLimit written on another. The limits are now updated under the lock with updateConf; value(forKey:) is untouched.
• Concurrent requests for the same URL with different imageIDs were coalesced into one TaskLoadImage, which stored the image only under the first request's keys, so the second ID missed the memory cache and downloaded again. The custom imageID is now part of TaskLoadImageKey (nil for default requests, so the common path hashes nothing extra); the data is still fetched once, by URL. This also fixes ImagePrefetcher deduplicating and stopping such requests under one key.
• containsCachedImage(caches: .disk) and containsData(for:) returned true for a request with .disableDiskCacheReads (or .reloadIgnoringCachedData) while cachedData and cachedImage returned nil for it: they queried the data cache without the read guard the other methods have.
• storeCachedImage encoded the image before checking whether there is a data cache or whether the request allows disk writes, so with the default configuration (no data cache) every call paid a synchronous encode and threw the result away. The data cache and the write option are now checked first.