Use compile-time autotune for the Qwen3.5 CUDA export on PyTorch 2.14 (#23140)

### Summary

`test-model-cuda-e2e-Qwen3.5-35B-A3B-HQQ-INT4-quantized-int4-tile-packed`
(and the Qwen3.5 `benchmark-cuda` job) have failed on every main run
since the PyTorch 2.14 pin bump (#22501).

The Qwen3.5 export set `autotune_at_compile_time=OFF`. That was added
with the 2.13 pin (#20967) to avoid a 2.13-only compile-time autotune
issue. On PyTorch 2.14, OFF selects AOTInductor's new lazy-compile path,
which breaks this model in two ways:

1. **Export crash.** Autotuning now runs a C++ JIT variant of the graph
(`GraphLowering._run_jit_variant_for_autotune`) directly on
`self.constants` and the real inputs. The low-memory CUDA export frees
KV-cache storage (`resize_(0)`) before compile. The #21803 rehydration
only hooks `CachingAutotuner.run`, so the null KV storage reaches the
kernels:
   ```
   _run_jit_variant_for_autotune -> compiled_fn(input_handles)
RuntimeError: CUDA driver error: an illegal memory access was
encountered
   ```
The logged error was only the secondary one, from `preserve_rng_state`
-> `set_rng_state`.
2. **Unrunnable artifact.** Even with the KV storage rehydrated, the
lazy AOT code loads every Triton kernel from a cubin file at runtime and
ignores `embed_kernel_binary`. ExecuTorch only packs the wrapper `.so`,
so the runner fails on the first execute with `CUDA driver error: file
not found`.

### Fix
Drop the `autotune_at_compile_time=OFF` override. On 2.14, AOTI defaults
it to on for accelerators, so kernels are autotuned at compile time on
generated example inputs and embedded in the `.so`, the same as the
other CUDA models. The obsolete 2.13 comment is removed too.

### Test plan
CI on this PR (earlier commits carried temporary diagnostics, which have
since been removed):
-
`test-model-cuda-e2e-Qwen3.5-35B-A3B-HQQ-INT4-quantized-int4-tile-packed`
passes:
  - export GPU peak 18682 MB (limit 21504 MB)
  - runner produces correct output, decode 113.8 tok/s
  - `Qwen3.5-MoE serving smoke passed`

Author: Gasoonjia

examples/models/qwen3_5_moe/export.py

@@ -1138,9 +1138,8 @@ def _export_cuda(model, config, args):
         "enable_dynamic_shape": True,
         "get_mutable_buffer_metadata": _mutable_buffer_metadata_json(model),
     }
-    # Avoid PyTorch 2.13's fused cast/reduction compile-time autotune path. Keep
-    # max autotuning enabled so prefill GEMMs use the CUDA backend's Triton-only
-    # GEMM lowering instead of emitting libtorch fallback kernels.
+    # Keep max autotuning enabled so prefill GEMMs use the CUDA backend's
+    # Triton-only GEMM lowering instead of emitting libtorch fallback kernels.
     et_prog = to_edge_transform_and_lower(
         {"decode": decode_ep, "prefill": prefill_ep},
         partitioner={
@@ -1151,7 +1150,6 @@ def _export_cuda(model, config, args):
                         CompileSpec("low_memory_mode", b"ON"),
                         CompileSpec("emulate_precision_casts", b"OFF"),
                         CompileSpec("max_autotune", b"ON"),
-                        CompileSpec("autotune_at_compile_time", b"OFF"),
                     ]
                 )
             ],
@@ -1162,7 +1160,6 @@ def _export_cuda(model, config, args):
                         CompileSpec("low_memory_mode", b"ON"),
                         CompileSpec("emulate_precision_casts", b"OFF"),
                         CompileSpec("max_autotune", b"ON"),
-                        CompileSpec("autotune_at_compile_time", b"OFF"),
                     ]
                 )
             ],