fix(payload-filter): resolve parameter filtering bugs for OpenAI-comp…

[zhchang]

…atible providers

Fix three critical bugs preventing parameter filtering from working correctly with OpenAI-compatible providers like Grok and DeepSeek:

Bug #1: Executor ordering - ApplyThinking re-adds filtered parameters

• Move ApplyPayloadConfigWithRoot() to run AFTER ApplyThinking() in openai_compat_executor
• Ensures filter has final authority over parameters sent to upstream provider
• Fixes: reasoning_effort re-injected after filter removes it for suffix models (e.g., grok-3(medium))

Bug router-for-me#2: Protocol mismatch - Responses API not recognized by filter rules

• Normalize "openai-response" protocol to "openai" in payloadModelRulesMatch()
• Filter rules with protocol: "openai" now match both Chat Completions and Responses API
• Fixes: Silent filter failures for Responses API requests

Bug router-for-me#3: Case-sensitive model matching

• Make matchModelPattern() case-insensitive by lowercasing both pattern and model
• Pattern "grok-*" now correctly matches "Grok-3", "GROK-3", etc.
• Improves robustness for real-world model name variations

Testing

• Added 19 comprehensive unit tests in payload_helpers_test.go
• Added 9 QA reproduction tests in test/reasoning_effort_filter_test.go
• All tests passing (28/28), zero regressions
• Verified end-to-end with real provider scenarios

Documentation

• Added FILTER_ARCHITECTURE.md documenting filter system design
• Provides reference for future maintenance and extensibility