fix: avoid conflicting token limits in summarization requests

[CBOSSX]

Closes #288.

Summary generation currently sends both max_tokens and max_completion_tokens, causing strict OpenAI-compatible endpoints to return HTTP 400. Select max_completion_tokens for OpenAI-compatible models and the common max_tokens option for native Claude/Anthropic models, so each provider receives its supported output budget.

The regression tests exercise real Eino request serialization against local HTTP fixtures for both classic and agentic paths. Both OpenAI cases fail before the fix with the reported 400 and pass afterward. Both native Claude cases retain the 4096-token summary budget instead of the model's 8192-token default. No live provider credentials are required.

Validation:

• go test ./internal/multiagent -run '^TestEinoSummarization(SendsOneTokenLimit|PreservesClaudeTokenLimit)$' -count=1 — failing before, passing after.
• go test ./internal/multiagent ./internal/openai ./internal/llm -count=1 — passed.
• go test ./... — passed.
• go vet ./internal/multiagent ./internal/openai ./internal/llm — passed.
• gofmt and git diff --check — passed.