CUDA: vectorize FP16 attention input preprocessing

[john-sp]

• Add an eight-elements-per-thread FP16 path for PE-dense attention-body preprocessing.
• Use aligned 128-bit loads and stores where channel boundaries permit.
• Retain the scalar implementation for FP32, non-dense positional encoding, and unsupported alignment.

Dispatch depends only on datatype, layout, and channel alignment; it contains no model, batch-size, device-name, or SM-count checks.

Performance

RTX 3090 Ti, BT4, batch 84:

• Master median: 23.732 us
• Optimized median: 14.246 us
• Kernel reduction: 40.0%
• Saving: approximately 9.5 us per evaluation

AI Acknowledgement: Written by Codex