[GGUF] Fix Gemma3 quantization support

This commit implements complete GGUF quantization support for Gemma3 models
with true Q4_0 compression, addressing gibberish output and enabling 50%
memory reduction.

Changes:
1. gguf_loader.py: Add gemma3_text -> gemma3 model type mapping
2. gemma3.py:
   - Add Gemma3 RMSNorm weight correction (-1.0 offset)
   - Fix qweight_type tensor shape (scalar -> [1])
   - Fix F16 embedding handling (no reshape needed)
   - Enable GGUF quantization in linear layers
   - Handle UninitializedParameter for GGUF layers

Key fixes:
- RMSNorm correction: Gemma3 uses (1+weight) convention but GGUF stores
  full values, requiring -1.0 subtraction
- F16 embeddings: GGUF raw data is already in PyTorch layout, preventing
  data corruption from unnecessary reshape operations
- qweight_type shape: GGUF layers expect shape [1] not scalar []

Tested on:
- 8 Gemma3 variants (1B-27B parameters)
- Both instruction-tuned and pretrained versions
- Q4_0 quantization format
- 100% success rate with coherent text generation

Fixes #14753, #15480

Signed-off-by: Luciano Martins <lucianommartins@users.noreply.github.com>

Author: lucianommartins

vllm/model_executor/model_loader/gguf_loader.py

@@ -63,6 +63,10 @@ def _get_gguf_weights_map(self, model_config: ModelConfig):
         # hack: ggufs have a different name than transformers
         if model_type == "cohere":
             model_type = "command-r"
+        if model_type == "gemma3_text":
+            # Gemma3 models use "gemma3_text" in HuggingFace but
+            # "gemma3" in GGUF architecture naming
+            model_type = "gemma3"
         if model_type in ("deepseek_v3", "deepseek_v2"):
             model_type = "deepseek2"
             # GGUF layer map assumes that we will have a merged expert weights