Apply one little-endian token bitmask to every row along the final axis.
Borrow logits; return an owned lazy result. Copy the mutable host bits now,
so a matcher can advance before GPU evaluation. Missing words reject tokens.
Preserve the additive-mask expression, including NaNs and signed zero.
Shape-specialized compilation handles bf16/f16/f32 and strided inputs.
Apply one little-endian token bitmask to every row along the final axis. Borrow logits; return an owned lazy result. Copy the mutable host bits now, so a matcher can advance before GPU evaluation. Missing words reject tokens. Preserve the additive-mask expression, including NaNs and signed zero. Shape-specialized compilation handles bf16/f16/f32 and strided inputs.