CUDA FA: run KV_max mask scan for all Q batch sizes

[ssam18]

The KV_max mask scan was gated behind Q->ne[1] >= 1024, so decode steps never ran it. This caused the flash attention vec kernel on Pascal GPUs to read past valid KV entries into uninitialized memory beyond the current context, triggering device-side exceptions above 24K tokens. Removing the batch-size guard fixes the crash with negligible overhead since the scan is a single small kernel launch. Fixes #22032

[JohannesGaessler]

This is the wrong fix. Those indices are solely for optimization, any out-of-bounds checks should be working regardless.