fix: address review feedback on auth inputs and dead base-action options

- Drop anthropic_api_key from repo workflows: only CLAUDE_CODE_OAUTH_TOKEN
  is configured as a secret, and upstream docs say to pick one auth path
- Replace the undeclared allowed_tools input in test workflows with
  claude_args --allowedTools (base-action only declares claude_args)
- Remove dead ClaudeOptions fields and INPUT_* reads that prepareRunConfig
  never consumed (mcp_config, allowed_tools, max_turns, system prompts, etc.)
- Update base-action README to document the actual inputs, with former
  dedicated inputs expressed as Claude CLI flags via claude_args

.github/workflows/test-custom-executables.yml

@@ -49,11 +49,10 @@ jobs:
         with:
           prompt: |
             List the files in the current directory starting with "package"
-          anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
           claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
           path_to_claude_code_executable: /home/runner/.local/bin/claude
           path_to_bun_executable: /home/runner/.bun/bin/bun
-          allowed_tools: "LS,Read"
+          claude_args: --allowedTools LS,Read
 
       - name: Verify custom executables worked
         run: |