Add application-level extension field and PrepareCall hook to AgentCall / AgentStreamCall

Currently AgentCall and AgentStreamCall only carry a fixed set of fields (sampling params, messages, tools, callbacks, etc.). There is no first-class way for application-layer code to:

1. Attach arbitrary, application-specific data to a call.
2. Run a one-shot preparation hook before the first model invocation, analogous to the internal agent.prepareCall.

Typical use cases for the hook: translating an application-defined context object into SystemPrompt / ProviderOptions / Messages, injecting tenant-scoped headers, resolving feature flags, hydrating RAG context, etc.

Start by locating AgentCall and AgentStreamCall, then read the internal agent.prepareCall flow to understand when preparation occurs. Define how application-specific data and a one-shot preparation hook fit both call types, with completion shown by supporting context translation, option or message injection, and other application-level preparation before the first model invocation.

Written by the indexing model from the issue text.