feat: add per-model and default request params

[ivanpashenko]

Summary

• add params to models() for provider request defaults per real model ID
• support default as either a model string or { model, ...requestDefaults }
• deep-merge configured defaults into outgoing LLM payloads
• let aliases and @default inherit their resolved model configuration
• preserve per-chat overrides

Example

models({
  default: {
    model: "openai/gpt-5.6-terra",
    reasoning: { max_tokens: 8000 }
  }
})

Validation

• npm test -- params_merge
• linked the local package in ~/.tune
• verified the object-form default sends reasoning.max_tokens: 8000 and completes a real Tune request