Normalize LLM usage telemetry through a typed adapter

Parent tracking issue: #4904

Desired Behavior

LLM telemetry should normalize provider-specific completion and Responses API usage objects once at a typed boundary, then compute metrics from one internal usage representation using direct attribute access. Provider aliases such as prompt/input tokens, completion/output tokens, token details, cache reads/writes, and hidden response metadata should not be rediscovered repeatedly with getattr.

Scope

• openhands-sdk/openhands/sdk/llm/utils/telemetry.py
• Focused internal usage/response adapter types and telemetry tests
• Dynamic-attribute baseline updates for this slice only

Acceptance Criteria

• A typed internal usage representation covers completion and Responses API token fields.
• Provider-shape normalization occurs once at the external response boundary.
• Sync and async telemetry paths reuse the same normalized implementation.
• Tests cover prompt/input, completion/output, reasoning, cache-read, cache-write, and missing-usage cases.
• Telemetry's forbidden dynamic-attribute baseline entries are removed and relevant tests/type checks pass.

This issue was created by an AI agent (OpenHands) on behalf of the user.

Start with openhands-sdk/openhands/sdk/llm/utils/telemetry.py and the existing telemetry tests. Trace both sync and async paths and identify where provider-specific usage shapes are currently inspected. Done means a shared typed usage representation, boundary normalization, coverage for the listed token and missing-usage cases, and passing relevant tests and type checks.