Use max example seq_len when exporting Qwen3.5 MoE

The previous example used T=2, which caused AOTI to compile the
chunk_gated_delta_rule kernel for a single chunk (NT=1). At runtime,
prompts longer than 64 tokens (requiring NT>1 chunks) failed with
"Error resizing tensor at input 0". Using max_seq_len-1 as the
example ensures AOTI generalizes intermediate buffer sizes for the
full sequence length range.

Comparison against original export (tq4_sdpa fused kernel)
on H100 (Qwen3.5-35B-A3B, HQQ-INT4, max_seq_len=4096, 5 runs median):

                Original (tq4_sdpa)  Baseline (Triton SDPA)
  Decode tok/s       68.4               61.7
  Prefill tok/s     275.7              378.2

Baseline prefill is 1.37x faster; decode is 0.90x (tq4_sdpa's fused
decode kernel is faster than the tiled Triton SDPA at L_q=1). The
split-K commit addresses the decode gap.

This PR was authored with the assistance of Claude

Author: digantdesai

examples/models/qwen3_5_moe/export.py

@@ -626,9 +626,14 @@ def _export_cuda(model, config, args):
     print("Decode export successful!")
 
     # --- Prefill method (T>=2, dynamic shape) ---
+    # Example T must equal max_seq_len-1 so AOTI compiles kernels (especially
+    # chunk_gated_delta_rule with CHUNK_SIZE=64) for the full range of sequence
+    # lengths. Smaller examples cause AOTI to bake in intermediate buffer sizes
+    # that reject longer prompts at runtime.
     print("Exporting prefill method...")
-    prefill_tokens = torch.tensor([[0, 1]], dtype=torch.long)
-    prefill_pos = torch.tensor([0, 1], dtype=torch.long)
+    example_prefill_len = config.max_seq_len - 1
+    prefill_tokens = torch.zeros((1, example_prefill_len), dtype=torch.long)
+    prefill_pos = torch.arange(example_prefill_len, dtype=torch.long)
     seq_dim = Dim("seq_len", min=2, max=config.max_seq_len - 1)
     prefill_dynamic_shapes = (
         {1: seq_dim},  # tokens