Add Gemma3 Model support for NvTensorRtRtx execution provider (microsoft#1520)

## Problem
Gemma3 models failed to build with NvTensorRtRtx execution provider due
to:
- Unsupported standalone RotaryEmbedding nodes
- Dual RoPE configuration requirements (local vs global attention
layers)
- Missing cos/sin cache inputs to GroupQueryAttention nodes

## Root Cause
1. NvTensorRtRtx doesn't support standalone RotaryEmbedding contrib ops
2. Gemma3 uses mixed attention pattern: 5/6 layers use local attention 
(sliding window + rope_local_base_freq), 1/6 layers use global attention
3. Previous implementation created incompatible RotaryEmbedding nodes
4. GroupQueryAttention expected cos/sin caches but they weren't provided

## Solution
- **Dual Cache System**: Create separate cos/sin caches for local vs
global layers
  - `cos_cache_global/sin_cache_global`
  - `cos_cache_local/sin_cache_local` (theta=rope_local_base_freq)
- **Skip RotaryEmbedding Nodes**: For NvTensorRtRtx, skip standalone
node creation
- **Internal RoPE Processing**: Set `use_rope_in_attn=True` to handle
RoPE inside GroupQueryAttention
- **Dynamic Layer Configuration**: Automatically select appropriate
cache/theta based on layer type using sliding_window_pattern

## Implementation Details
- Layer pattern: `(layer_id + 1) % sliding_window_pattern == 0` →
Global, else Local
- NvTensorRtRtx: RoPE handled internally by GroupQueryAttention with
proper cache inputs
- Other EPs: Standard RotaryEmbedding nodes with appropriate caches

## Testing
✅ Gemma3-4b-it model builds successfully with NvTensorRtRtx EP
✅ Maintains compatibility with other execution providers

@BLSharda @kunal-vaishnavi @baijumeswani

Author: anujj

src/python/py/models/builder.py

@@ -361,11 +361,7 @@ def make_attention_init(self):
             )
 
             # Some EPs don't support fusing rotary embeddings inside GQA yet
-            self.attention_attrs["use_rope_in_attn"] = (
-                self.ep not in ["dml", "webgpu"]
-                and not self.attention_attrs["q_norm"]
-                and not self.attention_attrs["k_norm"]
-            )
+            self.attention_attrs["use_rope_in_attn"] = self.ep not in ["dml", "webgpu"]
             if self.attention_attrs["use_rope_in_attn"]:
                 # GQA + Rot.Emb. does not require `position_ids` as input
                 self.input_names.remove("position_ids")