feat: improve test coverage and enforce variable gradient consistency

[Rakshitha-Ireddi]

Summary

This PR improves the reliability of the get_engine loading mechanism and enforces graph consistency in Variable.

Changes

1. Test Coverage: Added tests/test_get_engine.py which uses mocks to verify that get_engine correctly dispatches to OpenAI, Anthropic, Gemini, Cohere, and LiteLLM providers without requiring real API keys.
2. Core Refactor: Converted Variable.requires_grad to a property. The setter now validates that you cannot set requires_grad=False if any predecessors have requires_grad=True, preventing invalid computation graph states (addressing a TODO in variable.py).
3. Verification: Added tests/test_variable_grad.py to verify the new gradient invariant logic.

Motivation

• The get_engine function has complex dispatch logic that was previously untested.
• The gradient consistency check ensures users don't accidentally break the autograd graph.

Authors

• Ireddi Rakshitha
• Yaswanth Devavarapu