feat: add ContextLimit helper to modelinfo

[dgageot]

Model context-window handling was scattered across providers. Each provider either hardcoded a limit (Anthropic had anthropicContextLimit embedded in its package) or lacked a fallback when the models.dev store was unavailable, risking silent failures or degraded behavior in production environments.

This change centralizes context-window logic into a new ContextLimit(ctx, store, id, fallback) helper in pkg/modelinfo. It first attempts to source the window from the models.dev store with a bounded timeout, then falls back to a provided default. A new DefaultAnthropicContextLimit constant (200000) replaces Anthropic's hardcoded function, and all five provider packages now use modelinfo.ContextLimit consistently.

Additionally, LoadCaps now accepts a context.Context as its first parameter so callers can implement proper timeout behavior. This is wrapped internally with loadCapsTimeout to bound the underlying models.dev load operation, though callers should understand that the timeout does not cover contention from the shared store lock.

[docker-agent]

Assessment: 🟢 APPROVE

The ContextLimit helper is implemented correctly. The nil-store early-return, bounded context.WithTimeout, and <= 0 fallback guard all work as intended. The DefaultAnthropicContextLimit constant migration is type-safe (untyped integer constant implicitly converts to int64). All five provider convertDocument call-sites are updated consistently. The LoadCaps context threading is correct. Tests cover the happy path, unknown model, and nil-store cases.

No high or medium severity bugs were found in the changed code.