[Bug]: ruff-legacy pre-commit hook reports all baseline violations as regressions on Windows

Summary

On Windows, the ruff-legacy pre-commit hook reports every violation already
recorded in ruff-legacy-baseline.json as a new regression. It fails on files
that have not been touched at all, so the hook blocks every commit that
includes a legacy file and gives Windows contributors no way to distinguish a
real regression from the noise.

Reproduction

On a Windows checkout, with a clean working tree:

$ git status --porcelain          # empty
$ python scripts/legacy_utils.py lint-precommit \
      tensorrt_llm/serve/tool_parser/base_tool_parser.py \
      tests/unittest/llmapi/apps/test_tool_parsers.py

New violations (28 regressions vs baseline):
  tensorrt_llm\serve\tool_parser\base_tool_parser.py:46  D212  Multi-line docstring summary should start at the first line
  tensorrt_llm\serve\tool_parser\base_tool_parser.py:93  D205  1 blank line required between summary line and description
  ...
  tests\unittest\llmapi\apps\test_tool_parsers.py:2072  D202  No blank lines allowed after function docstring (found 1)

Both files are unmodified, and the reported counts match the baseline exactly:
base_tool_parser.py has D212: 6 in the baseline and six D212 violations in
the working tree.

Cause

normalize_path in scripts/legacy_utils.py builds the lookup key with
str(p.relative_to(repo_root)), which uses the platform separator:

https://github.com/NVIDIA/TensorRT-LLM/blob/main/scripts/legacy_utils.py#L124-L132

On Windows that yields tensorrt_llm\serve\tool_parser\base_tool_parser.py,
while ruff-legacy-baseline.json stores forward-slash keys. baseline.get(filepath, {})
in compare_against_baseline therefore always misses, every rule falls back to
a baseline count of 0, and every current violation counts as a regression.

The same normalized key feeds current_counts, checked_files,
identify_new_violations and the auto-fix report, so a single fix in
normalize_path covers all of them.

Suggested fix

Return a POSIX path, matching how the baseline is stored:

return p.relative_to(repo_root).as_posix()

This is a no-op on Linux and macOS, where as_posix() already matches str().

Environment

• Windows 11, Git Bash / PowerShell
• TensorRT-LLM main at f8c7f55b2b

Start in scripts/legacy_utils.py at normalize_path, then read how its result is used by compare_against_baseline and the related violation-reporting paths. Make the normalized key match the forward-slash paths in ruff-legacy-baseline.json, then rerun the provided Windows reproduction and confirm unchanged files no longer report baseline violations as regressions.