feat: filter None values from inputs before API requests

[zeke]

UPDATE: Moving this back to draft pending an internal discussion about whether this change is actually needed.

This PR filters out None-valued inputs from all prediction and training creation methods before making API requests.

Changes

• Add filter_none_values() utility function to recursively remove None values from dictionaries
• Update encode_json() and async_encode_json() to filter None values when processing dicts
• Apply filtering to deployments.predictions.create() and trainings.create() methods
• Add comprehensive test suite for None filtering functionality

Example

Before:

replicate.use("black-forest-labs/flux-schnell")(prompt="banana", seed=None)
# Would send: {"prompt": "banana", "seed": null}

After:

replicate.use("black-forest-labs/flux-schnell")(prompt="banana", seed=None)
# Now sends: {"prompt": "banana"}

This preserves the legacy behavior where None-valued inputs are automatically removed before making API requests, preventing potential API errors.

Fixes https://linear.app/replicate/issue/DP-737

[zeke]

Red herring. Closing.