Add MiniMax Anthropic-compatible provider support with model facts and regional endpoints

[octo-patch]

Reason: Add Anthropic-compatible provider support so the findings filter can route Claude API calls through a configurable provider endpoint with provider-supplied model identifiers and per-model facts.

Changes

• Add a provider registry (claudecode/providers.py) for Anthropic-compatible providers. It records the MiniMax provider with the MiniMax-M3 and MiniMax-M2.7 model identifiers and their per-model facts (context window, USD-per-million-token pricing, input modalities and thinking modes) plus global (https://api.minimax.io/anthropic) and China (https://api.minimaxi.com/anthropic) regional endpoint configuration.
• Wire the registry into ClaudeAPIClient and get_claude_api_client via optional provider and region parameters. When a provider is supplied, the Anthropic-compatible base URL is resolved from the registry for the chosen region (or the provider default region); an explicit base_url still takes precedence.
• Thread provider/region through FindingsFilter and initialize_findings_filter, reading CLAUDE_PROVIDER and CLAUDE_REGION environment variables.
• Add claude-base-url, claude-provider and claude-region action inputs in action.yml and document them (and the MiniMax model facts and regional endpoints) in the README.

Checks

• python3 -m py_compile on changed modules → ok.
• python3 -m pytest claudecode/ -q → 211 passed.

Notes

• The provider endpoint, regional endpoints and model identifiers/facts come from the provider configuration and the provider registry; no other provider values are hardcoded in product code. Tests exercise the regional endpoints and model facts to verify they are exposed correctly.

[octo-patch]

Updated MiniMax provider resolution so omitting claude-model or claude-base-url now uses MiniMax-M3 and the selected regional Anthropic endpoint for both Claude Code and findings filtering; invalid provider, model, or region values now fail fast. I ran the focused Python tests (83 passed), the remaining Python suite (218 passed with the existing CLI no-key integration assertion excluded), and Bun tests (11 passed).