Ownership of tsk_tree_t breaks in release builds.

[molpopgen]

Post-mortem for future reference:

In #300, we moved the Tree API to a Deref target.

The Tree itself still owns a tsk_tree_t on the stack.
However, in release builds, this model breaks.
The API type that is the Deref target takes a pointer
to the tsk_tree_t.
In release builds, that pointer is invalidated, presumably
due to final moves of all the bits into structs.

To get this right, we have to do what we do for tables,
which is to use malloc'd pointers to tsk_foo via MBox.

[molpopgen]

Checking out 0.9.0 -- all tests pass with --release.
Checking out 0.10.0 -- all tests pass with --release.

[molpopgen]

Whatever is going on happened in #300.
That is the PR that separated out the
tree interface as a deref target.

To discover this, I did a manual "git bisect";

• hard reset to the commit merging Separate ownership and API of Tree. #300
  • cargo --release test barfs
• Then git reset --hard HEAD~1 to get to one commit earlier
  • tests pass in release mode

[molpopgen]

For some insane reason, the low-level tree index
gets initialized to 0 in release mode but to -1
in debug mode. The latter is correct.