Harden DB-GPT service endpoints

[m1n9yu3]

Summary

This PR applies defensive hardening across several DB-GPT service endpoints and helpers:

• Tightens agent file download path handling and route protection.
• Restricts plugin hub updates to expected repository URLs.
• Replaces regex-heavy SQL comment stripping with a linear scanner.
• Adds route-level user dependencies to sensitive knowledge APIs.
• Simplifies PDF title matching to avoid pathological regex behavior.
• Enforces configured API keys consistently for Prompt APIs and v1 user dependencies.
• Parameterizes datasource configuration SQL and adds regression coverage.
• Scopes legacy resource file paths to conversation upload directories.
• Replaces v2 chat app stream regex parsing with linear JSON extraction.

Validation

• Added focused regression tests for authentication helper behavior.
• Added focused regression tests for datasource configuration DAO filtering and updates.
• Added focused regression tests for resource file path scoping.
• Added focused regression tests for v2 chat app stream parsing.
• Ran syntax checks for touched Python modules.
• Ran focused pytest coverage for the new and updated tests.

Notes

This public PR intentionally keeps the description high level. It does not include vulnerability reproduction details.

[chenliang15405]

The following issues still need to be resolved:

Issue 1: The file_client.py behavior change will break downstream callers. The original FileClient.read_file(file_key) directly opens any path, while the new code restricts it to the conv_uid upload directory — if a downstream caller passes an absolute path outside conv_uid, it will throw ValueError at runtime. All callers must be verified for compatibility.

Issue 2: The auth module default behavior is unclear. When API_KEYS is empty, do _validate_api_key and _require_user still allow all requests? If not, existing deployments without API key configuration will fail all authentication. The default behavior must be documented and backward compatibility ensured.

Issue 3: This PR contains 10 unrelated commits with an oversized impact scope. 16 files / 10 commits cover file downloads, hub URL validation, SQL comment stripping, knowledge base auth, PDF title matching, Prompt API Key, global API Key, datasource SQL parameterization, file path scoping, and SSE parsing — these changes have no correlation with each other, and the impact spans core/serve/app packages. If any single fix introduces a bug, it cannot be individually reverted.