Fix where_self argument order in attention mask construction

In tch-rs, `where_self(condition, other)` wraps `torch.where(condition,
self, other)` where `condition` must be a Bool tensor. The previous code
incorrectly passed `allow_mask` (Bool) as `self` and `zero` (Float) as
`condition`, which swaps the roles of the condition and value tensors.

The comment `// where(allow, 0, -inf)` makes the intended semantics
clear: return 0 where allow is true, -inf where false. The correct call
is `zero.where_self(&allow_mask, &neg_inf)` so that `allow_mask` (Bool)
is the condition, `zero` is the value where true, and `neg_inf` is the
value where false.

Signed-off-by: Michael Yuan <michael@secondstate.io>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: juntao

src/audio_encoder.rs

@@ -236,7 +236,7 @@ impl AudioEncoder {
             );
             // where(allow, 0, -inf)
             let mask = Tensor::from_tch(
-                allow_mask.as_tch().where_self(&zero.into_tch(), &neg_inf.into_tch())
+                zero.into_tch().where_self(&allow_mask.as_tch(), &neg_inf.into_tch())
             );
             Some(mask)
         }