fix segfault on watch test, #58

[yjh0502]

The segfault (#58) happens when watch outlives database/cluster. It happens befause TrxWatch does not keep any refcount to database/cluster. I fixed the issue by making all FdbFuture live with either Transaction or Database.

I found another possible problem while fixing the issue, which is caused by drop order of Rust. Rust drops struct field from top to bottom[1]. If we define Cluster prior to Database, cluster will be dropped before Database, which might cause a problem. valgrind does not complain for this case but I changed the field order to just make sure.

[1] https://github.com/rust-lang/rfcs/blob/master/text/1857-stabilize-drop-order.md#tuples-structs-and-enum-variants

[bluejekyll]

Rust drops struct field from top to bottom[1].

I don't believe that linked RFC can be relied on, and it doesn't look like there is a path forward on this: rust-lang/rfcs#1857

---

Edit: I just reread that, and it was merged... it looks like it's a non-change, meaning they will be keeping drop order. I completely missed this, and thought something happened related to struct packing.

[bluejekyll]

Ignore my previous comments, my understanding of drop order was outdated.

This looks good, if a little confusing. We basically will now have two Futures, one scoped to Database, and one to Transaction, as I understand it. Which seems logical.