Add adaptive thinking support to the computer-use demo

[lg-ant]

Newer Claude models (Opus 4.6 / Sonnet 4.6 and later) reason adaptively — the model decides how much to think, steered by an effort level — instead of using a manual thinking token budget. This adds adaptive thinking support to the computer-use reference implementation:

• Each model config declares which thinking modes it supports (adaptive / extended / off), and the sidebar offers only those modes: an effort slider for adaptive, the existing budget input for extended.
• Adds Claude Opus 4.6, Sonnet 4.6, Opus 4.7, and Claude Opus 4.8 to the model map, and defaults the demo to claude-opus-4-8 with adaptive thinking at medium effort.
• Model IDs resolve by longest known prefix, so undated aliases and Bedrock/Vertex forms of a known family get the right capabilities; unrecognized IDs keep all modes selectable.
• Older models keep their existing behavior (extended thinking with a token budget).
• Refreshes the README callout to reflect the new default model.

Verification: 51 unit tests pass, including new cases covering the request body per thinking mode and the per-model UI options; ruff and pyright clean.