fix(#1743): retry primary model on transient errors even without fallback models

[aheritier]

Summary

Fixes #1743

getEffectiveRetries returned 0 when no fallback models were configured, meaning retryable errors (5xx, timeouts) got zero retries. This caused Anthropic streaming Internal server error to immediately surface as "all models failed" instead of being retried with backoff.

Root Cause

The getEffectiveRetries function conflated "no fallback models to fall back to" with "no need to retry the same model":

// Before: returns 0 when no fallback models configured
if retries == 0 && len(a.FallbackModels()) > 0 {
    return DefaultFallbackRetries
}
return retries // returns 0

With maxAttempts = 1 + 0 = 1, even correctly-classified retryable errors got exactly one shot.

Fix

Changed getEffectiveRetries to always return DefaultFallbackRetries (2 retries = 3 total attempts) when no explicit retry count is configured, regardless of whether fallback models exist:

// After: always returns DefaultFallbackRetries when not explicitly configured
if retries == 0 {
    return DefaultFallbackRetries
}
return retries

Tests

• Updated TestGetEffectiveRetries to expect DefaultFallbackRetries for agents without fallback models
• Added TestPrimaryRetriesWithoutFallbackModels — regression test reproducing the exact scenario (Anthropic streaming Internal server error, no fallback models, verifies retry and recovery)
• Added isRetryableModelError test case for the Anthropic streaming error format