This has proven to be intractible: various attempts to eliminate have failed, so detect at runtime and cease the traversal (and do the remaining ops outside a traverse).

Note: we can still deadlock on traversal vs transaction corner cases.
We handle transactions as single operation, which it logically is.

If we do the trace inside the lock, we show the actual sequence number when the transaction started.

Also changes first member of transaction to have valid start_group field,
and fix outdated comment.

(Invasive: means passing full op array all the way through).

Generalizes traverse in traverse fix from rusty@rustcorp.com.au-20090629073630-3eduhyypx2tp6u80

We now use the "group_len" field for both transactions and traversals.

Make sort_deps more efficient, and also only alter order when necessary.  This means by default we run in serial number order, only going outside when we detect a dependency.
Maintain trace file order in original sort, so sort_deps doesn't mess it up.

We still need serial numbers: sort_deps can have multiple solutions for a single key, but these may deadlock with the ordering requirements of other keys.  By sticking close to the actual order (ie. serial order), we minimize the chance of this happening.

In my tdbtorture -n 4 example trace, this reduces from 14493 to 3210 dependencies, but doesn't make any measurable improvement in the time.  Still, it's simple to do and might make a difference for larger sets.

Also, using a single structure makes talloc_free convenient for when we do optimization.

Still deadlocks in one case, due to spurious dependencies inside traversals.  See next commit.

realloc() has that horrible overloaded free semantics.

By faking them out to length 1, we might go oob.  Just fake the malloc.