[Feature Request] Add Native Sparse Attention (NSA) support

Is your feature request related to a problem? Please describe.

Training and inference with long sequences (32K+) using dense attention is prohibitively expensive. TE currently offers sliding window attention as the only sparse alternative, but this uses a fixed local pattern that loses long-range dependencies.

Describe the solution you'd like

Add support for Native Sparse Attention (NSA) from the paper Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention. NSA dynamically selects important tokens via learned compression and top-k block selection, maintaining long-range dependencies while being hardware-efficient.

Describe alternatives you've considered

Using the standalone Triton implementation (native-sparse-attention-triton) - Works but doesn't integrate with TE's FP8 quantization, CP, or fused kernels. Requires maintaining separate code paths.

Additional context

Start by reading the linked Native Sparse Attention paper and the standalone native-sparse-attention-triton implementation. The issue names no TransformerEngine files or tests; done would require an agreed integration of NSA with FP8 quantization, context parallelism, and fused kernels.