A Go binding for inotify

This is a fork of the since-deleted golang.org/x/exp/inotify package. It breaks the API in an attempt to fix some reliability problems:

1. inotify watches are attached to inodes rather than file paths. A single inode could have multiple file paths, or change its file path between when it is watched and when events are generated.

2. AddWatch now returns a *Watch representing the watch descriptor. Watches can be compared by pointer equality, as only one watch struct will be used for each watch descriptor. For convenience, the watch stores the path it was created for.

3. There is no attempt to automatically use IN_MASK_ADD. The previous behaviour was unreliable, so you'll need to specify it manually if desired.

4. Event now includes a Watch field for its associated watch (or nil for events not associated with a watch like IN_Q_OVERFLOW). The Name field is no longer prefixed with the watch's path.

5. The Watcher.Error channel has been removed. Errors are now reported when Close is called.

6. The Watcher's mutex is now used to guard all access to the watches map.

7. When RemoveWatch is called, we don't completely forget about the watch until we receive an associated IN_IGNORED event. This way any queued events for the old watch can be correctly decoded.

8. Reads from the inotify instance are performed through an *os.File. This should take advantage of the Go runtime's IO polling system and stop the readEvents goroutine from tieing up an OS thread.

There is still a potential concurrency issue with the approach I've taken. Imagine this sequence of events:

1. user calls RemoveWatch for watch descriptor N
2. the Read call in the readEvents goroutine returns with events for watch descriptor N
3. user calls AddWatch, which happens to reuse watch descriptor N
4. the readEvents goroutine acquires the lock to decode the events it received

With this sequence of events, it would be unclear which watch the event belonged to. According to this LKML post, watch descriptors are allocated sequentially, and can only be reused after they wrap at INT_MAX. So this is unlikely to be a problem in practice.