feat(session): expose maxOutputTokens on llm adapters

[uchouT]

Summary

• Add maxOutputTokens?: number to AnthropicAdapterOptions and OpenAICompatibleOptions so consumers can configure the per-request output cap once at adapter creation time
• New fallback chain for the request max_tokens: completeOptions.maxTokens (per-call override) > options.maxOutputTokens (adapter default) > 4096 (built-in safety net) — preserves existing behaviour when neither is set

Why

Both adapters previously hard-coded ?? 4096 for the request max_tokens. This is too low for non-trivial outputs: long tool-call argument JSONs (e.g. multi-item results) and long synthesis/integration outputs hit the cap and get truncated. The downstream parsers then throw on the unterminated JSON. Without a way for embedders to raise the cap at adapter level, every consumer had to either patch the source or pass maxTokens on every single call site.

Note: Anthropic's messages.create API requires max_tokens, so the adapter must always send a value — maxOutputTokens only adjusts which value is sent, not whether one is sent.

Test plan

• Unit tests for fallback precedence on both adapters (default 4096 / options.maxOutputTokens / completeOptions.maxTokens wins)
• Stream path uses the same value as complete path (anthropic)
• All existing session/core/devtools tests still pass (146 + 285 + 15)