[Examples] add tfla op and tfla-optimized

[Old-cpu]

Summary

Description

This PR adds an optimized implementation of the Tiled Flash Linear Attention (TFLA) kernel based on the NeurIPS 2025 paper "Tiled Flash Linear Attention: More Efficient Linear RNN and xLSTM Kernels",demonstrating significant performance improvements over the baseline versions. The optimized TFLA kernel achieves approximately 21× speedup compared to the baseline TFLA implementation, and approximately 3× speedup compared to the fused GQA Attention kernel (next-gqa-attention-fusion.mlir).

Hardware Configuration

• CPU: Intel Xeon Silver 4114 @ 2.20GHz
• Cores: 2 sockets × 10 cores = 20 physical cores (40 logical threads)
• Architecture: x86_64 with AVX-512 support
• OpenMP threads: 48

Background

The TFLA kernel provides a fair comparison with Grouped Query Attention (GQA) under identical configurations:

• Batch size: 1
• Query heads: 12
• KV groups: 2 (each group serves 6 heads via GQA)
• Sequence length: 1 (single-token inference)
• Hidden dimension per head: 128
• KV cache sequence length: 1024 per groupt.

Expected Performance

• Describe how reviewers can verify and test this change. Include command-line instructions if applicable.
  make next-tfla-run and make next-tfla-optimized-run

Checklist

• The code builds successfully
• Existing tests pass
• New tests are added where appropriate
• Code follows the project coding style
• Documentation is updated if needed