Consider including parameter type schemas in `Template` user message

[eb8680]

Templates face an issue similar to that outlined for Tools in #699

Templates carry both input and output types, but the LLM API currently only receives the schema of the output type (which is used for structured output generation). The input arguments to a Template call are encoded and interpolated into the user message in call_user, so the LLM sees their values but does not see their types, even when the types carry useful contextual information that complements what's in the prompt template.

We should therefore consider injecting the type schemas into the message generated by call_user upon first calling a Template.