(refactor): replace legacy Anthropic constants

[gold-silver-copper]

Summary

This PR updates rig-core's direct Anthropic provider to the current Claude model families and removes the old public Anthropic constants.

It also fixes a shared streaming test helper bug that surfaced once Anthropic started emitting short pre-tool assistant text before the final tool-backed answer.

What changed

• Removed the old Anthropic model constants from rig-core
• Added only the current family constants:
  • CLAUDE_OPUS_4_6
  • CLAUDE_SONNET_4_6
  • CLAUDE_HAIKU_4_5
• Updated rig-core examples, tests, and docs to use the new constants
• Refreshed Anthropic default max token resolution for the current 4.x families
• Switched Anthropic reasoning tests from thinking.enabled + budget_tokens to thinking.type = "adaptive"
• Fixed the shared reasoning streaming helper to compare FinalResponse.response() with the final assistant turn only, excluding pre-tool assistant chatter

Breaking changes

This PR intentionally removes the previous Anthropic constants with no alias or deprecation layer.

Removed public API items:

• anthropic::completion::CLAUDE_4_OPUS
• anthropic::completion::CLAUDE_4_SONNET
• anthropic::completion::CLAUDE_3_5_HAIKU

Replacement migration path:

• CLAUDE_4_OPUS -> CLAUDE_OPUS_4_6
• CLAUDE_4_SONNET -> CLAUDE_SONNET_4_6
• CLAUDE_3_5_HAIKU -> CLAUDE_HAIKU_4_5