[GGUF] Revert GGUF WA for GPU

[sammysun0711]

Details:
Revert GGUF Reader WA for OV GPU plugin: #2110

Ticket:
CVS-169891

[Wovchena]

The same test that was skipped for Mac fails for Linux. Maybe resolving it for Linux fixes it for Mac as well...

[sammysun0711]

The same test that was skipped for Mac fails for Linux. Maybe resolving it for Linux fixes it for Mac as well...

Although same test test_full_gguf_pipeline failed on Mac and Linux. But the issue are different.

• Skipped Mac test failed with segment fault with 2nd test model
  

  openvino.genai/tests/python_tests/data/models.py

  Lines 80 to 81 in 6a620fb

  	"gguf_model_id": "Qwen/Qwen2.5-0.5B-Instruct-GGUF",
  	"gguf_filename": "qwen2.5-0.5b-instruct-q4_0.gguf"

• Current Linux test case failed with 3rd test model due to output mismatch w/ HF model
  

  openvino.genai/tests/python_tests/data/models.py

  Lines 84 to 85 in 6a620fb

  	"gguf_model_id": "sammysun0711/tiny-random-deepseek-distill-qwen-gguf",
  	"gguf_filename": "tiny-random-deepseek-distill-qwen_q8_0.gguf"

Will take a look for proper fix the test case.

[Copilot]

Pull Request Overview

This PR reverts previous workarounds (WA) for GPU plugin issues in the GGUF model handling code. The changes remove GPU-specific fixes that were implemented to address accuracy and compilation issues on MTL/LNL GPU platforms.

• Removes shared embedding parameter and logic from language model creation
• Reverts dynamic quantization group size from 0 (disabled) back to 64 (enabled)
• Removes zero point array modification workaround for Q4_0 weights

Reviewed Changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
src/cpp/src/gguf_utils/gguf_modeling.cpp	Removes shared_embedding parameter, reverts dynamic quantization settings, and cleans up GPU-related comments
src/cpp/src/gguf_utils/building_blocks.hpp	Updates make_lm_head function signature to remove shared_embedding parameter
src/cpp/src/gguf_utils/building_blocks.cpp	Removes shared embedding logic and zero point array modification workaround

[Copilot]

[nitpick] The logic structure has changed after removing the shared_embedding condition, but the original fallback logic (using embeddings_node when key + ".weight" doesn't exist) is preserved. Consider adding a comment to clarify this fallback behavior for future maintainers.

[sammysun0711]

@Wovchena, updated GGUF CI tests and all test passed, local GPU test passed, can we merge this PR?