First commit to interface ensemble launcher and chemgraph

[harikrishna1410]

This pull request is a preliminary effort to integrate chemgraph with ensemble launcher. As a first step,

1. I added a el_mcp_tools.py, which is a replication of mcp_tools.py tools.
2. run_el_mcp.py: A runner script to start the mcp server

This is not expected to be the final version. I wanted to start the discussion here.

[tdpham2]

@harikrishna1410 Thanks for working on this! I have a few comments:

• Could you retarget this PR to dev and rebase from the latest dev instead of main?
• Architecture. I thought about this and I think there are two options for supporting both EL and Parsl:

Option 1: Parallel MCP Servers. Each backend gets its own MCP server file(s):

src/chemgraph/mcp/
---mcp_tools.py # Local execution (no parallelism)
---mace_mcp_parsl.py # MACE via Parsl
---graspa_mcp_parsl.py # gRASPA via Parsl
---mace_mcp_el.py # MACE via EnsembleLauncher
---graspa_mcp_el.py # gRASPA via EnsembleLauncher
---ase_mcp_el.py # General ASE via EL (NEW)

Option 2: Execution backend abstraction. A common interface so a single ensemble MCP server can dispatch to any backend:

class ExecutionBackend(ABC):
    """Abstraction for dispatching compute tasks."""
class EnsembleLauncherBackend(ExecutionBackend):
    """EL"""
class ParslBackend(ExecutionBackend):
    """Parsl"""

And a single ensemble MCP server:

backend = load_backend(os.getenv("EXECUTION_BACKEND", "local"))
@mcp.tool(name="run_mace_ensemble")
async def run_mace_ensemble(params: MaceEnsembleInput) -> dict:
    futures = []
    for struct in params.structures:
        fut = backend.submit(run_mace_core, {"params": struct})
        futures.append(fut)
    results = await backend.gather(futures)
    return {"status": "success", "results": results}

I think Option 2 is more scalable and future-proof. However, if you want to start with Option 1 to get something working first, that's fine. I can start refactoring toward Option 2 in a follow-up once both backends are functional.

[harikrishna1410]

I like option 2. this is mainly because tools can be added independently of the execution backend.