fix(llm): add conditional max_tokens parameter naming per provider

[ajianaz]

Problem

LLM API providers use different parameter names for max output tokens:

Provider	Parameter
OpenAI (legacy)	max_tokens
OpenAI (new)	max_completion_tokens
Anthropic	max_tokens
Google Gemini / Vertex	max_output_tokens
Some proxies	map max_tokens → max_output_tokens and then reject

cora-cli previously hardcoded max_tokens everywhere. When a provider doesn't support it, we get HTTP 400: Unsupported parameter: max_output_tokens.

Solution

Add max_tokens_param to the llm section in .cora.yaml. Supports:

• auto (default): detect from provider name.
  • gemini / google / vertex → max_output_tokens
  • Everything else → max_tokens
• max_tokens, max_output_tokens, max_completion_tokens: explicit override

Config example

llm:
  max_tokens_param: auto       # default — detect from provider
  # max_tokens_param: max_output_tokens  # explicit override

Changes

• src/config/schema.rs: Added max_tokens_param: Option<String> to LlmSection and max_tokens_param: String to Config (default: "auto")
• src/engine/types.rs: Added max_tokens_param: String to LLMConfig (default: "max_tokens")
• src/config/loader.rs: Added resolve_max_tokens_param() for auto-detection; resolved value passed to LLMConfig
• src/engine/llm.rs: Replaced hardcoded "max_tokens" JSON key with dynamic config.max_tokens_param in both streaming and non-streaming request paths
• src/commands/config_cmd.rs: Display max_tokens_param in cora config output

Tests added

• test_resolve_max_tokens_param_auto_* — auto-detection for known providers (gemini, google, vertex, openai, anthropic)
• test_resolve_max_tokens_param_explicit_override — explicit value wins over auto
• test_chat_request_uses_max_output_tokens — verify JSON key name is correct
• test_chat_request_uses_max_tokens_default — verify default key name
• Schema merge tests for max_tokens_param propagation
• YAML roundtrip test for LlmSection with max_tokens_param

All 621 tests pass. cargo fmt --all -- --check clean.