Idea: Can we cache the result of `typeToTypeNode`?

A large percentage of declaration emit time (and a small percentage of check time) is spent in typeToTypeNode and tracing indicates that it is called on some types many times (hundreds or thousands in a sufficiently large project). I suspect, but have been able to demonstrate, that those requiring a import() qualification are particularly expensive. Could we cache the result? Two clear obstacles:

1. Nodes are mutated during emit, so we'd need to clone anything retrieved from the cache. I'm reasonably confident that cloning will be cheaper than recomputing.
2. typeToTypeNode takes a complex context argument. Possible workarounds include including the context hash in the cache key and storing the cached value on the context.