Fix DeepCompile fallback handling for scalar outputs (#8149)
When DeepCompile lowered the graph-safe RNG wrapper around this
operator, the fallback handler treated every non-`None` output as an
Inductor buffer for never-reuse registration and called `get_name()`
unconditionally.
The scalar output therefore raised `AttributeError: 'int' object has no
attribute 'get_name'` before the first compiled forward could run (See
the following definition of `_scaled_dot_product_flash_attention`).
```
aten::_scaled_dot_product_flash_attention(
Tensor query,
Tensor key,
Tensor value,
float dropout_p=0.0,
bool is_causal=False,
bool return_debug_mask=False,
*,
float? scale=None
)
->
(
Tensor output,
Tensor logsumexp,
Tensor cum_seq_q,
Tensor cum_seq_k,
SymInt max_q,
SymInt max_k,
Tensor rng_state,
Tensor unused,
Tensor debug_attn_mask
```
This PR limits "never-reuse" registration to Inductor IR outputs. This
is applied to tensor buffers, while scalar and `None` leaves pass
through without tensor-only method calls.
Signed-off-by: Masahiro Tanaka <mtanaka@anyscale.com>
