LLMConfig.max_tokens not respected by OpenAI-based clients

Bug Description

When configuring max_tokens in LLMConfig, the setting is ignored by OpenAI-based clients (OpenAIClient and AzureOpenAILLMClient), causing token limit errors even when a higher limit is configured.

Additionally, there are currently two confusing ways to set max_tokens for OpenAI clients:

Through LLMConfig(max_tokens=...)
Through the client constructor parameter OpenAIClient(max_tokens=...)

This dual configuration is confusing because:

It's unclear which one takes precedence
Users naturally expect the config object to be the authoritative source
The constructor parameter silently overrides the config setting
Other clients (Gemini, Anthropic) don't have this confusing dual interface

What I was trying to do

Configure a higher token limit for LLM responses to avoid truncation:

from graphiti_core.llm_client.config import LLMConfig
from graphiti_core.llm_client.openai_client import OpenAIClient

config = LLMConfig(
    api_key="...",
    model="gpt-4",
    max_tokens=16000  # Set higher limit
)

client = OpenAIClient(config=config)

What I expected to happen

The client should respect the max_tokens=16000 setting and allow responses up to 16,000 tokens, similar to how other configuration options work.

What actually happened

The client still uses the default 8192 token limit, resulting in errors like: Output length exceeded max tokens 8192: Could not parse response content as the length limit was reached

Root Cause

In BaseOpenAIClient.__init__(), the constructor parameter max_tokens=DEFAULT_MAX_TOKENS always overrides config.max_tokens, even when explicitly set by the user:

# Current problematic code:
def __init__(self, config: LLMConfig | None = None, cache: bool = False, max_tokens: int = DEFAULT_MAX_TOKENS):
    # ...
    super().__init__(config, cache)
    self.max_tokens = max_tokens  # This overwrites config.max_tokens!

Code sample that demonstrates the issue

from graphiti_core.llm_client.config import LLMConfig
from graphiti_core.llm_client.openai_client import OpenAIClient

config = LLMConfig(max_tokens=16000)
client = OpenAIClient(config=config)

print(f"Config max_tokens: {config.max_tokens}")  # Shows: 16000
print(f"Client max_tokens: {client.max_tokens}")  # Shows: 8192 (wrong!)

# User expects: client.max_tokens == 16000
# Actual result: client.max_tokens == 8192

Affected Clients

OpenAIClient
AzureOpenAILLMClient

Other clients handle this correctly:

GeminiClient - respects config and has clear precedence rules
AnthropicClient - respects config settings
GroqClient - respects config settings

Proposed Solution

The fix should:

Respect config.max_tokens when explicitly set (different from default)
Use constructor parameter as fallback when config has default value
Maintain backward compatibility for existing code
Clear precedence rules: Config takes priority over constructor parameter

# Proposed fix:
if config.max_tokens != DEFAULT_MAX_TOKENS:
    self.max_tokens = config.max_tokens  # Use config when explicitly set
else:
    self.max_tokens = max_tokens  # Use constructor parameter as fallback