feat(core): add token_budget parameter to limit retry token usage

[Jwrede]

Summary

Refs #2056

Adds an optional token_budget parameter that stops retries when cumulative token usage exceeds the specified limit. This mitigates the retry amplification vector where adversarial or malformed LLM outputs cause repeated validation failures, leading to unbounded context growth and cost amplification.

Usage

client.chat.completions.create(
    response_model=MyModel,
    max_retries=10,
    token_budget=5000,  # stop retrying after 5000 total tokens used
)

When the budget is exceeded between retry attempts, the retry loop stops and raises InstructorRetryException (same as when max_retries is exhausted).

Implementation

• Adds token_budget and usage_container parameters to initialize_retrying()
• Uses a closure-based tenacity stop condition that checks cumulative usage (supports both OpenAI total_tokens and Anthropic input_tokens + output_tokens)
• Extracts token_budget from kwargs in both retry_sync and retry_async (same pattern as existing timeout)
• No breaking changes -- parameter is optional, behavior is unchanged when not set

Test plan

• test_token_budget_stops_retries -- verifies retries stop before max_retries when budget exceeded
• test_token_budget_not_set_retries_normally -- verifies default behavior unchanged
• test_token_budget_success_before_limit -- verifies successful responses return normally
• Existing retry tests pass (test_retry_json_mode.py)
• ruff check + format clean