Rodrigo Moraes created some special loaders for Jinja2 under GAE, see here. It's not bytecode caching but it precompiles all templates to Python so you avoid the Jinja2 parsing overhead.
Note that (from this GAE page):
compiled application code is cached
for rapid responses to web requests