fix(contrib): improve adopt docstring example to show meaningful convergence and add doctest coverage

[wahid18-maqs]

Summary

This PR improves the adopt() docstring example to show meaningful optimizer convergence and adds doctest coverage to keep the example executable.

Changes

• Update the adopt() docstring example to:
  • Use learning_rate=0.1 instead of 0.003 so the optimizer makes visible progress
  • Use jax.value_and_grad instead of jax.grad to follow the idiomatic optax pattern
  • Update expected outputs to match the corrected example
• Add optax/contrib/_adopt_test.py with a doctest.DocTestSuite wrapper to execute the example as part of the test suite

Why

The previous example used learning_rate=0.003 which produced no visible decrease in the objective function across 5 steps — the loss stayed at 14.0 throughout. This misled users into thinking ADOPT was not working or that stagnation was expected behavior.

The updated example uses a more appropriate learning rate that shows the optimizer actually making progress, giving users a realistic expectation of what ADOPT does.

Adding doctest coverage also ensures the example remains valid and synchronized with the implementation going forward.