Add MiniMax as first-class LLM provider

[octo-patch]

Summary

Adds MiniMax as a first-class LLM provider alongside OpenAI, with full support for PDF VLM scraping, webpage analysis, agentic chunking, and structured extraction.

Changes

• thepipe/provider.py (new): Provider abstraction with ProviderPreset dataclass, create_provider_client() factory, auto-detection via MINIMAX_API_KEY, temperature clamping, and think-tag stripping
• thepipe/__init__.py: Added --provider and --api-key CLI flags for provider selection (backwards-compatible with existing --openai-* flags)
• thepipe/chunker.py: chunk_agentic() now falls back from .beta.chat.completions.parse() to json_object mode when structured output is not supported (e.g. MiniMax)
• README.md: MiniMax setup guide with Python/CLI examples, models table, and provider selection docs
• tests/test_provider.py: 31 unit tests + 3 integration tests

MiniMax Models

Model	Context	Notes
MiniMax-M2.7	1M tokens	Latest, recommended
MiniMax-M2.7-highspeed	1M tokens	Faster inference
MiniMax-M2.5	204K tokens	Previous generation
MiniMax-M2.5-highspeed	204K tokens	Fast inference

Test plan

• 31 unit tests passing (provider presets, factory, temperature clamping, think-tag stripping, agentic chunking fallback)
• 3 integration tests passing with real MiniMax API (chat completion, JSON mode, PDF scraping)
• Existing test suite unaffected (17 passing, 2 pre-existing failures unrelated to changes)