fix: Handling of sync and async requests

[casperlehmann]

Summary

This PR fixes Fakesnow's compatibility with both Node.js and Python Snowflake SDKs by implementing the correct response format and query result polling behavior.

Changes

1. Add status code for successful queries (ff43a93)

• Added code: "0" field to POST response to explicitly signal "results ready"
• This field must be set in the response, since the clients are not doing null handling:
  • Without this field, SDKs interpreted responses differently:
    • Node.js SDK: Assumed async query → attempted GET request (decision logic)
    • Python SDK: Treated as request failure → retried with new POST requests (polling logic)
• Now both SDKs correctly use inline results without additional requests.
• Returning code: "0" does mean that we are essentially never serving async responses. But since I'm still not sure how to make that decision programmatically, I have left it like this for now.

2. Add required response fields (a711a9c)

• Added fields expected by both SDKs: queryId, sqlText, returned, total, etc.
• Ensures response structure matches SDK expectations for both Node.js (rowset) and Python (rowsetBase64)

3. Add async GET endpoint support (13ad8de)

• Implemented get_cached_query_result() function to retrieve cached query results
• Mounted /queries/{query_id}/result endpoint for async polling behavior
• Allows SDKs to fetch results via GET when code: "333333" (in progress) or code: "333334" (async execution) is returned
• Added results cache (conn.results_cache) with LRU eviction (max 50 entries) to store query results by queryId
• Note: This polling endpoint was hit on every single query sent by the async Node.js SDK. After adding the "code" field in the query_request response, this is no longer the case. Since we added a default value (code: "0" -- completed) to query_request, the request will never enter the polling state, and this GET endpoint will no longer get hit. But I am leaving it in, since it is already working and, ideally, the query_request should be able to return in-progress values for code as well (perhaps by setting a timeout in on the query execution).

4. Add SafeJSONResponse for datetime serialization (88e0bfd)

• Created custom response class to handle non-JSON-serializable types like datetime
• Prevents serialization errors when query results contain temporal data
• Uses custom JSON encoder that converts problematic types to strings

5. Store results cache in tuple format with rowtype (a16fd35, 63e73fe)

• Internal cache only: Results cache (conn.results_cache) is never exposed externally - only used for result_scan() and get_results_from_sfqid()
• Changed cache format from dictionary to 6-tuple: (arrow_table, rowcount, last_sql, last_params, last_transformed, rowtype)
• Rowtype is required for GET /queries/{query_id}/result endpoint to generate rowsetBase64 (Arrow IPC format needed by Python SDK)
• Cache uses OrderedDict for LRU eviction (max 50 entries) to prevent unbounded memory growth
• Fixed tuple format consistency: both server.py and cursor.py now store 6-tuple with rowtype

6. Add support for persisting database in server mode

Allow clients the option to set FAKESNOW_DB_PATH when running in server mode, in order to persist data on disk.

Technical Details

The changes implement the private SDK API (/queries/v1/query-request) used by Node.js and Python Snowflake connectors, not the public Snowflake SQL API. Key differences:

• Success code: "0" (not "090001" from public API)
• Data fields: rowset/rowsetBase64 (not data array from public API)
• Async codes: "333333" (in progress), "333334" (detached) - defined in Node.js SDK and Python SDK

See also: Snowflake SQL API Reference for comparison with public API behavior.