[Bugfix] Fix compressed-tensors models failing to load with transformers backend

[mgoin]

Purpose

The transformers backend's WeightsMapper uses a broad prefix mapping ("" -> "model.") to transform weight paths. This mapper is also applied to quantization config targets via apply_vllm_mapper. However, compressed-tensors targets can be module class names (e.g., "Linear") or regex patterns (e.g., "re:.*proj"), not just layer paths.

When "Linear" is transformed to "model.Linear", the target matching fails because the module class name check looks for "Linear" in "model.Linear" (substring match), which fails.

The fix filters which targets get transformed: only layer paths (containing . and not starting with re:) are mapped. Class names and regex patterns are preserved.

Test Plan

# Previously failing, now passing - compressed-tensors with transformers backend
vllm serve RedHatAI/Qwen3-0.6B-FP8-BLOCK --model-impl transformers

# Verify native backend still works
vllm serve RedHatAI/Qwen3-0.6B-FP8-BLOCK

# Verify fp8 (non-compressed-tensors) still works with transformers
vllm serve Qwen/Qwen3-0.6B-FP8 --model-impl transformers

Test Result

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[gemini-code-assist]

Code Review

This pull request correctly identifies and addresses an issue where the WeightsMapper was incorrectly transforming non-path targets in compressed-tensors quantization configurations. The approach of conditionally applying the mapping is sound. However, I've identified a subtle bug in the implementation of the helper functions _apply_dict and _apply_list. They use a truthiness check that would incorrectly filter out targets mapping to an empty string, potentially leading to silent misconfigurations. My review includes a suggested fix to ensure correct behavior by using an explicit is not None check, aligning with the original WeightsMapper logic.

[Isotr0py]

LGTM

[mgoin]

cc @eldarkurtic to fix your issue reported offline