Replace hardcoded Gemma config dicts with HF config cache entries

Move all Gemma 1, 2, and 3 models to _HF_CONFIG_CACHE using their
native HF config classes (GemmaConfig, Gemma2Config, Gemma3TextConfig,
Gemma3Config). Add generic architecture handlers for GemmaForCausalLM,
Gemma2ForCausalLM, Gemma3ForCausalLM, and Gemma3ForConditionalGeneration
that read from the HF config objects, eliminating ~460 lines of
hardcoded dicts and the name-based architecture detection.

Fixes google/gemma-2b incorrectly getting Gemma2ForCausalLM architecture
(now correctly gets GemmaForCausalLM from cache). Gemma 2 attn_types now
exactly match n_layers instead of having extra unused entries.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: claude

tests/unit/test_convert_hf_model_config.py

@@ -403,10 +403,7 @@ def test_gemma_2b(self):
                 "n_key_value_heads": 1,
                 "gated_mlp": True,
                 "final_rms": True,
-                # NOTE: "google/gemma-2b" contains "gemma-2" so the architecture
-                # detection assigns "Gemma2ForCausalLM" even though this is a Gemma 1 model.
-                # The config values are still correct (matched by model name prefix).
-                "original_architecture": "Gemma2ForCausalLM",
+                "original_architecture": "GemmaForCausalLM",
             },
         )
 
@@ -457,7 +454,7 @@ def test_gemma_2_2b(self):
                 "n_key_value_heads": 4,
                 "window_size": 4096,
                 "use_local_attn": True,
-                "attn_types": ["global", "local"] * 21,
+                "attn_types": ["global", "local"] * 13,  # 26 layers
                 "attn_scores_soft_cap": 50.0,
                 "output_logits_soft_cap": 30.0,
                 "gated_mlp": True,
@@ -488,7 +485,7 @@ def test_gemma_2_9b(self):
                 "n_key_value_heads": 8,
                 "window_size": 4096,
                 "use_local_attn": True,
-                "attn_types": ["global", "local"] * 21,
+                "attn_types": ["global", "local"] * 21,  # 42 layers
                 "attn_scores_soft_cap": 50.0,
                 "output_logits_soft_cap": 30.0,
                 "gated_mlp": True,
@@ -520,7 +517,7 @@ def test_gemma_2_27b(self):
                 "n_key_value_heads": 16,
                 "window_size": 4096,
                 "use_local_attn": True,
-                "attn_types": ["global", "local"] * 23,
+                "attn_types": ["global", "local"] * 23,  # 46 layers
                 "attn_scores_soft_cap": 50.0,
                 "output_logits_soft_cap": 30.0,
                 "gated_mlp": True,