fix: implement lazy client creation in replicate.use()

[zeke]

Problem

replicate.use() was creating a client immediately, causing errors when no API token was available at call time. This prevented usage in Cog pipelines where tokens are provided through cog.current_scope() at runtime.

Solution

Modified Function and AsyncFunction classes to support lazy client creation by always accepting client classes and instantiating them on demand.

Changes

• _module_client.py: Pass Replicate and AsyncReplicate classes directly instead of factory functions
• lib/_predictions_use.py: Simplified to always use Type[Client] approach:
  • Removed Union[Client, Type[Client]] in favor of just Type[Client]
  • Updated constructors to store _client_class instead of mixed types
  • Simplified _client property to always instantiate from class
  • Updated all use() overloads to accept class types only
  • Use issubclass() for async client detection
• Added simple test demonstrating the fix

Result

✅ replicate.use("model") works even when no token is initially available
✅ Client created lazily when model is actually called
✅ Token retrieved from current cog scope at call time
✅ Simplified, consistent approach using only class types
✅ Clean type inference without complex Union handling

Fixes the original error:

replicate.ReplicateError: The bearer_token client option must be set either by passing bearer_token to the client or by setting the REPLICATE_API_TOKEN environment variable