Investigate codegen for global state in rust bindings

Right now, in crates/bindings/src/lib.rs we cache a row buffer in a thread_local. A comment claims that this gets optimized away, but I don't see that happening. I see calls for getting and using the local key still that are definitely not inlined and definitely not optimized out.

Removing that overhead while providing a safe interface to the bindings just takes some doing. I think if we can require code using this buffer to not be reentrant, we can recover that promise of optimizing down to just a normal wasm global lookup, while still giving access to a &mut Vec reference. Other strategies may trade eg RefCell counting for ease of implementation. I haven't measured yet.

It probably isn't low hanging fruit yet. But as the new memory architecture comes online, I expect to see more of this sort of overhead show up.

cf https://github.com/clockworklabs/SpacetimeDB/pull/452/files

cc @gefjon

Start in crates/bindings/src/lib.rs and compare the current thread_local buffer handling with the linked pull request 452. Measure the overhead of local-key access and determine whether a safe non-reentrant approach or another strategy meets the optimization goal. Done means the overhead is measured and a concrete implementation direction is established.