fix(ext): normalize Semantic Kernel token usage

[Thota Shashank (thotashashank302)]

Why are these changes needed?

SKChatCompletionAdapter reports zero token usage for Anthropic responses because it only reads OpenAI's prompt_tokens and completion_tokens attributes. For example, Claude usage with input_tokens=147 and output_tokens=1024 becomes RequestUsage(0, 0), affecting the returned usage, cumulative counters, and logged usage.

Normalize usage in both create() and create_stream(). Accept SDK usage objects and mappings: Semantic Kernel's Anthropic connector returns an object for ordinary responses and a dictionary for streaming responses. Prefer OpenAI field names when present, including explicit zero values, and fall back to Anthropic names. Missing usage retains the existing zero-count behavior.

Add 16 offline regression cases covering both request paths, Anthropic and OpenAI usage objects/dictionaries, field precedence, missing/null usage, accumulated totals across two calls, and usage events.

Validation on macOS with Python 3.12.14, using uv, lockfile-pinned Semantic Kernel/Anthropic dependencies and check tools:

• New regression cases on the original implementation: 6 failed, 10 passed.
• Complete Semantic Kernel adapter test module after the fix: 26 passed.
• Ruff lint and formatting, mypy, and Pyright on both changed files: passed.
• autogen-ext wheel build and git diff --check: passed.

The full workspace/optional-provider test matrix and documentation build were not run locally. Tests use mocked responses; no live model APIs were called. Implementation and local validation were assisted by OpenAI Codex.

Related issue number

Fixes #5464

Checks

• Reviewed documentation impact; this fixes existing token accounting without changing public APIs.
• Added tests corresponding to the changes.
• All GitHub checks have passed (pending CI).