Support per-row path in named_credential_request_col

[rkang-sf]

Summary

named_credential_request_col could only vary the request body per row; the url came from the HTTPRequest template and was the same for every row. This adds an optional per-row path Column that is appended to request.url, so each row can target a different endpoint of the same Named Credential (e.g. pick a model, per-entity GET lookups, geocoding):

REQUEST = HTTPRequestBuilder().set_url("callout:gemini").set_method(HTTPMethod.POST).build()
path = when(col("premium"), lit("/gemini-2.5-pro")).otherwise(lit("/gemini-2.5-flash"))
df = df.withColumn("resp", named_credential_request_col(REQUEST, body=col("payload"), path=path))

Join rule (rows can never switch Named Credential):

• path starts with / or ? → appended as is: callout:gemini + /v1 → callout:gemini/v1; + ?k=a → callout:gemini?k=a
• any other path → joined with /: callout:gemini + 2/v1 → callout:gemini/2/v1 (not callout:gemini2/v1)
• null or empty path → request.url
• path omitted → exactly the previous behavior. named_credential_request_col only forwards path when given, so SparkNamedCredential implementations that don't accept it yet keep working for existing callers.

Changes

• SparkNamedCredential.request_col: new optional path keyword; docstrings updated.
• DefaultSparkNamedCredential.request_col (local dev): with path, the UDF is applied over (body, path); each row uses a copy of the template whose url is join_callout_path(request.url, path).
• named_credential_request_col: new optional path argument.
• Separate Lint fixes commit: make lint currently fails on main (pretty-format-toml on pyproject.toml, black on tests/spark/test_column_hints_deferred.py). Formatting only; the reformatted pyproject.toml parses to identical data.

The Data Cloud runtime's built-in named_credential_callout UDF already takes the url per row, so no server-side change is needed; the runtime-side join lives in the internal SDK with the same rule.

Testing

• Unit tests: path forwarding, per-row path join, separator rule (incl. 2/v1, @evil.example.com), null/empty fallback, path without body. Full suite: 625 passed.
• pre-commit run --all-files --hook-stage manual (same as make lint): all hooks pass.
• Local Spark smoke test with a real DataFrame over /gemini-2.5-pro, ?key=a, 2/v1, @evil.example.com, "", NULL: urls joined as above, empty/null fall back to callout:gemini, omitting path is unchanged.