We missed a place where we invoke container while the async runtime is kept alive, in postblit.

Having a global runtime without a clear guarantee around lifetime makes it very easy to miss this. We can fix this in two ways, both eliminate the use of a global static runtime:

• Pass the runtime around the codebase giving us precise control of when to drop / re-init a new one
• Make block_on always construct a single use runtime that is dropped before returning the completed task

I've chosen the latter for simplicity, even if its less efficient. Its impossible to misuse. The first approach still runs risk of someone calling into container from some nested function that isn't aware an active runtime is held across its call site. This means that unless we pass ownership of runtime into and out of every function, we can easily lose track and hit this bug again.