Choose more consistent commit behavior

Right now we have two commit events: one for wlr_surface, one for wlr_output. Each of these structs have a state struct with a pending and current field.

When commit is emitted, I'd expect these assumptions to be true:

1. Users should have access to the committed state in the current state
2. Users should be able to figure out what changed in this commit (how?)
3. The pending state should be empty (it's just been applied)

Questions:

• Should we allow access to the previous state? Note that we don't want to retain locks on buffers longer than necessary.
• pending.committed is a bitmask of fields changed since last commit. What should current.committed mean? Right now it seems like it's the union of all commits' pending.committed.

Right now (1) is true, however:

• wlr_output clears the pending state after commit is emitted, breaking (3). The pending state is used by users to figure out what changed (2). wlr_output provides no access to the previous state.
• wlr_surface has a previous state, however previous.buffer is an exception and is always set to NULL. (3) is true. Also, the previous state seems to be only used by wlr_surface itself. Users have no way to figure out what changed (2).

It would be nice to decide on a consistent behavior.