feat: add FallbackLLM for built-in provider failover

[Oxygen56]

Description

Implements automatic failover across multiple LLM providers with configurable retries and error classification.

Features

• FallbackLLM class wraps a list of LLM instances in priority order
• Error classification: transient errors (timeouts, 429s, 5xx, connection errors) trigger automatic failover to the next provider; permanent errors (4xx auth, bad requests, content policy violations) fail immediately
• Configurable retries: retry_attempts parameter controls how many times each provider is retried before falling back
• Observability: all fallback events are logged via Python's standard logging module
• Full endpoint coverage: supports chat, complete, stream_chat, stream_complete, and all async variants

Example Usage

from llama_index.llms.openai import OpenAI
from llama_index.llms.anthropic import Anthropic
from llama_index.core.llms.fallback import FallbackLLM

openai_llm = OpenAI(model="gpt-4o")
anthropic_llm = Anthropic(model="claude-sonnet-4-20250514")

llm = FallbackLLM(
    llms=[openai_llm, anthropic_llm],
    retry_attempts=2,
)
response = llm.complete("Hello, world!")

Files Changed

• llama-index-core/llama_index/core/llms/fallback.py - New FallbackLLM implementation
• llama-index-core/llama_index/core/llms/init.py - Add FallbackLLM to exports
• llama-index-core/tests/llms/test_fallback_llm.py - Comprehensive test suite (32 tests)

Closes #19631