Fix: run mcp-server-hello-world in stateless HTTP mode

[noramxiao]

Problem

mcp-server-hello-world builds its FastMCP app with the default stateful HTTP transport (mcp_server.http_app()). Stateful mode requires the client to carry an mcp-session-id header across requests. The Databricks Assistant does not send that header, so every POST creates a new session and the server rejects the follow-up request with HTTP 400 — even though the same server works fine in the Playground (which does maintain the session). See #141 for the reporter's trace.

Change

mcp_app = mcp_server.http_app(stateless_http=True)

Stateless mode makes each request self-contained, so clients that don't track a session id work. It also avoids session-affinity issues on horizontally scaled Databricks Apps where consecutive requests can land on different replicas. The server's tools (health, get_current_user) hold no per-session state, so there's no functional downside.

Fixes #141

[noramxiao]

Integration tests now run and pass with stateless_http=True:

tests/test_integration_server.py::test_list_tools PASSED   [ 50%]
tests/test_integration_server.py::test_call_tools PASSED   [100%]
======================== 2 passed, 4 warnings in 13.92s ========================

list_tools() and tool invocation both succeed in stateless mode. Still worth a manual Databricks Assistant smoke-test before merge, since that client (no mcp-session-id header) is the scenario the fix targets and isn't covered by these tests.

[noramxiao]

Direct before/after verification of the fix

Reproduced the exact #141 failure mode locally — a session-less request (what the Databricks Assistant sends, since it doesn't carry an mcp-session-id header):

POST /mcp   {"jsonrpc":"2.0","id":1,"method":"tools/list"}     # no mcp-session-id header

Server	Result
Stateful (http_app(), before fix)	HTTP 400 — {"error":{"code":-32600,"message":"Bad Request: Missing session ID"}}
Stateless (http_app(stateless_http=True), this PR)	HTTP 200 — returns the full tools/list result

The 400 is precisely the error reported in #141. With stateless_http=True the same session-less request succeeds, confirming session-less clients (the Assistant) can now list/call tools.

Integration tests still pass (2 passed).

[noramxiao]

Verified on a live Databricks Apps deployment (e2-dogfood staging)

Deployed this branch as a Databricks App and sent session-less requests (no mcp-session-id header) directly to <app-url>/mcp, mirroring the Databricks Assistant client from #141:

Request (no session id)	Result
tools/list	HTTP 200 — full tools list returned
tools/call -> health	HTTP 200 — {"status":"healthy","message":"Custom MCP Server is healthy and connected to Databricks Apps."}

Same requests against the unfixed stateful build return HTTP 400 "Missing session ID". The fix resolves #141 end-to-end on real Apps infra, not just locally.