Enable use_sdpa_with_kv_cache in the qwen3_5 example config (pytorch#22390)

Fixes pytorch#22045.

`use_sdpa_with_kv_cache: False` in this config costs roughly 2x decode.
@JakeStevens confirmed in the issue there is no known reason to keep it
off, and the lfm2 config — the same hybrid attention/conv architecture —
already sets it `True`.

Measured on Qwen3.5-2B, XNNPACK, 8da4w + 8-bit embedding, the same
checkpoint exported twice with only this line different (Mac arm64,
token-by-token prefill then greedy decode, same prompt, 32 new tokens):

| | decode |
|---|---|
| `use_sdpa_with_kv_cache: False` | 8.20 tok/s |
| `use_sdpa_with_kv_cache: True` | **16.64 tok/s** |

That it is the same model rather than a faster different one: three of
four prompts decode token-identically (the fourth diverges inside
`<think>`, a near-tie either way), first-step logits corr 0.995 with the
same top-1, and at 338 tokens of context corr 0.990 with the same top-1
and top-2. File size is unchanged (1483 MB both).


cc @mergennachin @iseeyuan @lucylq @helunwencser @tarun292 @kimishpatel
@jackzhxng

Author: john-rocky

examples/models/qwen3_5/config/qwen3_5_xnnpack_fp32.yaml

@@ -7,7 +7,7 @@ base:
 
 model:
   use_kv_cache: True
-  use_sdpa_with_kv_cache: False
+  use_sdpa_with_kv_cache: True
   enable_dynamic_shape: False
   dtype_override: fp32