fix(python): preserve max_output_tokens in ModelLimits

[HDMowri]

Summary

• Preserve max_output_tokens on the public Python ModelLimits type.
• Keep the field during model metadata deserialization and serialization.
• Preserve compatibility with existing positional ModelLimits construction.

Context

The generated RPC ModelCapabilitiesLimits already carries
max_output_tokens, but CopilotClient.list_models() converts model metadata
through the public ModelLimits.from_dict(), which previously discarded the
field.

The TypeScript public ModelCapabilities type used by listModels() already
includes max_output_tokens. This follows the same schema/public-type parity
principle addressed in #2569, although that PR was motivated by the TypeScript
session-override path while this change fixes Python model-metadata
deserialization.

Compatibility

ModelLimits already supports positional construction as:

ModelLimits(max_prompt_tokens, max_context_window_tokens, vision)

The new field is keyword-only so existing positional arguments retain their
meaning. This follows the compatibility approach used by the SDK's generated
Python dataclasses when fields are added after an established positional API.

Fixes #2798

Testing

• verifies max_output_tokens deserialization and serialization
• verifies existing positional ModelLimits construction remains unchanged
• pytest test_client.py — 154 passed
• pytest test_legacy_parameters.py — 7 passed
• ruff format --check copilot test_client.py
• ruff check copilot test_client.py
• ty check copilot