fix: activation memory config crashes get_default() with OpenAI

[anatolykoptev]

Summary

Two bugs in default_config.py that crash get_default() when enable_activation_memory=True:

Bug 1: act_mem injected into MOSConfig (extra=forbid)

get_default_config() adds act_mem dict to the config passed to MOSConfig(**config_dict), but MOSConfig has no act_mem field and inherits extra="forbid" from BaseConfig:

pydantic_core.ValidationError: 1 validation error for MOSConfig
act_mem
  Extra inputs are not permitted [type=extra_forbidden]

Bug 2: KV cache with openai backend

get_default_cube_config() hardcodes extractor_llm backend to "openai" for KV cache memory, but KVCacheMemoryConfig validator requires huggingface/huggingface_singleton/vllm (KV cache extracts internal attention tensors via build_kv_cache, which needs local model access):

ConfigurationError: KVCacheMemoryConfig requires extractor_llm backend
to be 'huggingface' or 'huggingface_singleton', got 'openai'

Reproduction

from memos.mem_os.utils.default_config import get_default

# This crashes with ValidationError
config, cube = get_default(
    openai_api_key="sk-...",
    enable_activation_memory=True,
)

Fix

• Remove act_mem from get_default_config() — the enable_activation_memory bool flag is sufficient for MOSConfig; act_mem config belongs in MemCube config only.
• In get_default_cube_config(), require explicit activation_memory_backend kwarg (e.g. "huggingface") instead of hardcoding "openai". Skip act_mem gracefully when no compatible local backend is provided.

Test plan

• get_default(openai_api_key="sk-...", enable_activation_memory=True) no longer crashes
• get_default(openai_api_key="sk-...", enable_activation_memory=True, activation_memory_backend="huggingface", activation_memory_llm_config={...}) creates valid act_mem config
• get_default(openai_api_key="sk-...") continues to work as before