Add Vast.ai Mem0 Qwen3.5 launcher

[jinuk0211]

What changed

• add a local OpenAI-compatible Mem0 preset for Qwen/Qwen3.5-4B
• add a minimal benchmark dependency set that stays compatible with a recent vLLM runtime
• add a Vast.ai launcher that creates the environment, starts chat and embedding servers, waits for health checks, generates the runtime YAML, and runs the benchmark
• make LongMemEval the default benchmark and support overriding its sample cap at runtime
• document smoke/full runs, GPU placement, and supported environment overrides
• enforce LF line endings for shell scripts

Why

The existing Mem0 preset assumes already-running local endpoints and does not provide a reproducible Vast.ai bootstrap path. Qwen3.5 also requires a newer serving stack than the repository-wide pinned Transformers dependency, so the launcher keeps vLLM setup separate from the broad all-method requirements.

User impact

Users can clone this branch on a CUDA-enabled Vast.ai instance and start a one-query LongMemEval smoke run with a single script invocation. MAX_QUERIES=0 runs every query in the configured five-sample preset, while MAX_TEST_SAMPLES can expand that sample limit.

Validation

• staged Bash script passes bash -n
• script is committed with executable mode and LF line endings
• agent and LongMemEval YAML parse and required fields were asserted
• runtime agent/dataset YAML generation was exercised, including the sample-limit override
• pip install --dry-run --ignore-installed -r requirements-mem0-vast.txt resolves successfully
• git diff --cached --check passes

GPU model loading was not executed on the Windows development host because no local CUDA/WSL runtime is available.