fix: add tool count validation with env-configurable MAX_TOOLS

[lireno]

Summary

Fixes aipotheosis-labs#337 — prevents cryptic OpenAI API errors when too many functions are selected (e.g., enabling all Gmail functions in the playground).

This is a cleaner, improved version of PR aipotheosis-labs#604 that adds env-var configurability, better error messaging, and comprehensive tests.

What changed

Core changes

• config.py: Added MAX_TOOLS = 64 default, overridable via SERVER_MAX_TOOLS env var
• exceptions.py: New ToolCountExceeded exception with actionable 400 response message
• routes/agent.py: Early tool-count validation in /chat route before any API call
• agent/prompt.py: Defensive warning log in openai_chat_stream for other call-sites
• tests/routes/agent/test_chat_tool_validation.py: 11 test cases

Improvements over PR aipotheosis-labs#604

1. Env-var configurability: MAX_TOOLS is configurable via SERVER_MAX_TOOLS, not a hardcoded constant
2. Better error messages: Shows actual tool count and limit, plus actionable guidance to reduce functions
3. Comprehensive tests: 11 test cases covering exception shape, config defaults, boundary conditions (at-limit, over-limit, under-limit), custom limits, and zero-tools edge case
4. Proper exception class: Uses ACIException subclass pattern consistent with the codebase

Why this fix works

Validates tool count before making the API call. Returns a 400 with a clear message telling users to reduce selected functions. The default limit of 64 is conservative — the actual OpenAI limit is closer to 128 or ~200k tokens of tool definitions, but this provides a safe margin.

Risk assessment

• Low risk: validation-only change
• No behavioral change for requests under the limit
• Existing requests that would fail now fail earlier with a useful message
• Backward compatible

Closes aipotheosis-labs#337