[Gluon] Add 2CTA block-scaled matmul example with cuBLAS comparison (triton-lang#9697)

High-performance 2CTA warp-specialized block-scaled MMA. Two CTAs
cooperate per output tile, sharing operands to increase arithmetic
intensity and reduce the per-CTA SMEM footprint.
```
block-scale-matmul-mxfp8-mxfp8:
     MNK     1cta (TFLOPS)     2cta (TFLOPS)         2cta/1cta   cublas (TFLOPS)       2cta/cublas
    8192            2525.9            2895.0              1.15            2894.5              1.00
   16384            2409.3            2755.0              1.14            2647.7              1.04
   32768            2468.7            2632.9              1.07            2587.0              1.02

block-scale-matmul-nvfp4-nvfp4:
     MNK     1cta (TFLOPS)     2cta (TFLOPS)         2cta/1cta   cublas (TFLOPS)       2cta/cublas
    8192            4781.9            5730.0              1.20            5589.8              1.03
   16384            4837.0            5562.0              1.15            4313.6              1.29
   32768            4723.3            5362.4              1.14            4933.4              1.09

block-scale-matmul-mxfp8-mxfp4:
     MNK     1cta (TFLOPS)     2cta (TFLOPS)         2cta/1cta
    8192            2738.9            3149.8              1.15
   16384            2735.8            2930.4              1.07
   32768            2632.8            2773.1              1.05

block-scale-matmul-mxfp4-mxfp4:
     MNK     1cta (TFLOPS)     2cta (TFLOPS)         2cta/1cta
    8192            4956.5            5819.2              1.17
   16384            5196.6            5581.2              1.07
   32768            4862.2            5511.0              1.13
```

Author: adstraw