[codex] fix anthropic thinking max_tokens floor

[clifton]

Summary

This PR fixes issue #26 where Anthropic requests with Claude extended thinking could fail with:

Invalid request: max_tokens must be greater than thinking.budget_tokens

User impact

When users enabled thinking_level on Claude 4.x models without manually setting max_tokens, rstructor sent max_tokens=1024 by default while thinking budgets can be 1024/2048/4096/8192. That could produce immediate API errors for Minimal, Low, Medium, and High thinking levels.

Root cause

The Anthropic request builders hard-coded max_tokens to self.config.max_tokens.unwrap_or(1024) and did not enforce Anthropic's requirement that thinking.budget_tokens < max_tokens.

Anthropic docs specify this constraint in the extended thinking API docs:
https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking#the-thinking-parameter

Fix

• Added shared helper effective_max_tokens(configured_max_tokens, thinking_config) in src/backend/anthropic.rs.
• Enforced max_tokens > thinking.budget_tokens whenever thinking is enabled.
• Applied this logic to both:
  • structured output path (materialize_internal)
  • raw generation path (generate_with_metadata)
• Added a warning log when max_tokens is auto-adjusted.
• Updated thinking_level docs to state this auto-adjust behavior.

Tests

Added new unit tests in src/backend/anthropic.rs covering:

• default behavior without thinking
• configured behavior without thinking
• auto-adjust when default max tokens is too low
• auto-adjust when configured max tokens equals budget
• preserving configured value when already valid
• saturation edge case for extreme budget values

Validation

• cargo fmt --all
• cargo clippy --all-targets --all-features -- -D warnings
• cargo test --all-features

Closes #26.