Fix loading of quantized BigCode models

[eldarkurtic]

Loading of GPTBigCodeForCausalLM is broken due to the missing prefixes and incorrect treating of scales for QKV layers (i.e. c_attn.weight_scale, which are merged into a single matrix). Before this PR quantized models would produce garbage output.
After this PR quantized models are matching the BF16 baseline output.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[gemini-code-assist]

Code Review

This pull request aims to fix the loading of quantized BigCode models by adding necessary prefixes to layer names and correcting the handling of QKV layer scales. While the prefix additions seem correct, the change in how c_attn.weight_scale is loaded appears to be problematic and may lead to incorrect behavior or errors. The default weight loading logic for fused QKV layers seems to only update the scale for the 'query' component, leaving 'key' and 'value' scales at their default values, which is likely incorrect.

[mgoin]

LGTM, thank you!