feat(core): tolerant JSON parsing for LLM output (repair_json, loads_robust)

[Bartok9]

Problem

LLM responses very often emit JSON that is almost valid but breaks json.loads:

• wrapped in a Markdown code fence (```json … ```),
• containing // line comments or /* … */ block comments,
• containing trailing commas before } or ].

These defects are a recurring source of parse failures in the agent / output-parsing
paths. Today callers must hand-roll cleanup (and the existing _format_json_str
only handles newlines/tabs).

Solution

Two small, stdlib-only helpers added to dbgpt.util.json_utils:

• repair_json(text) — best-effort cleanup that, in order, strips a surrounding
  code fence, removes // and /* */ comments, and removes trailing commas before
  }/]. All transformations preserve the contents of double-quoted strings
  (escape-aware), so commas/slashes inside string values are never touched.
• loads_robust(text, **kwargs) — tries strict json.loads first and only falls
  back to parsing repair_json(text) if that raises JSONDecodeError. Well-formed
  input parses identically to json.loads; **kwargs are forwarded.

Tests

Added 12 tests in packages/dbgpt-core/src/dbgpt/util/tests/test_json_utils.py
covering trailing commas (object/array), fenced blocks (with/without language tag),
line/block comments, a combined case, string-content preservation, valid-JSON
passthrough, and the unrepairable-input raise path.

$ python -m pytest dbgpt/util/tests/test_json_utils.py -q
17 passed in 0.15s

Compatibility

• No breaking changes — purely additive; existing functions untouched.
• Pure standard library (no new dependencies).
• loads_robust only diverges from json.loads when strict parsing already fails.