Add chainable loading for Predict modules

[sushmanthreddy]

Currently, loading a predictor after optimization requires two separate lines:

predictor = dspy.Predict("question -> answer")
predictor.load(saved_path)

This PR adds support for chainable loading to improve code ergonomics:

predictor = dspy.Predict("question -> answer").load(saved_path, return_self=True)

Changes

• Modified load_state() and _load_state_legacy() to return self for chaining
• Added a new return_self parameter to load() method
• Added comprehensive tests for both traditional and chainable loading patterns
• Maintained backward compatibility with existing code by defaulting return_self=False

Testing

Added two new test cases:

• test_chainable_load: Verifies both traditional and chainable loading methods
• test_load_state_chaining: Ensures state loading supports chaining in both normal and legacy modes

Backward Compatibility

The changes are fully backward compatible:

• Existing code using load() without return_self continues to work as before
• New code can opt into chainable loading by setting return_self=True

Closes #1738

[isaacbmiller]

Nice thanks for writing this!