`TorrentSpec.Storage` was not honored when calling `Client.AddTorrentSpec`
and the configured `cfg.DefaultStorage` was always used. Now if you construct
your `TorrentSpec` you can specify any `StorageImpl`
Also, the most common use case for custom storage being per-torrent paths for
FileStorage, this adds a `pathMaker` function to the File implementation that
allows customization, along with the default (always use base path) and my use
case (which seemed common enough from the Gitter chat) which is infohash based
subdirectories.
All Public methods have not changed signature, but 1 private method did, hence
the test update.
This allows lots of behaviour to be baked into the new Client, Torrent and Piece wrappers, rather than duplicating (badly) them in all the backend implementations.
Torrent.updatePiecePriorities is always followed by Torrent.piecePriorityChanged, so move it into the former function. Also drop the separate mutex for Reader, and add a missing Reader.posChanged call.
It wasn't possible to move Client.WriteStatus to an external package to verify it doesn't depend on same-package access to data because then it can't be used in the tests, and it's extremely useful there. So I've settled for not locking the Client, and trying to use all the public methods. It's a work in progress.
Failing to get metainfo was resetting Torrent.metadataBytes, which caused arriving metadata data to appear to belong to nonexistent pieces, despite passing the check that we'd in fact requested them from the sending connection. This was unnecessarily noisy. Instead return an error if we got all the data, but couldn't set the Torrent's info bytes, that should propagate out and cause the connection that contributed the final missing piece to be dropped. This will at least provide some improved resistance to unfriendly behaviour on the network.
