Precompute shader stages / fix render pipeline

Rather than determining the amount and size of intermediate shader stages ahead of time, we easily can (and should) precompute this information. This would accomplish a number of goals:

1. We could let the output size dynamically match the input size after hooks (a vf_libplacebo requirement)
2. We could solve problems like #132 (and improve stuff like plane merging) by providing up-front information on exactly when and where scaling is required, and in which direction
3. If the API is designed correctly, we could also figure out a way to provide pass and timing data for debugging/informational purposes