Add a nvfp4 gemv example

[vickiw973]

No description provided.

[vickiw973]

the host contains many torch tensor creation thus the benchmark costs will be longer as the cuda event will count costs of the entire custom_kernel.

[vickiw973]

The tests can pass with the latest CuTe DSL release

pip3 install --upgrade nvidia-cutlass-dsl --pre

For NVFP4 GEMV (using FFMA to simulate the computation logic)

reference-kernels/problems/nvidia/nvfp4_gemv> python3 eval.py test task.yml 
compile: start
compile: pass
test-count: 10
test.0.spec: m: 128; k: 256; l: 1; seed: 1111
test.0.status: pass
test.1.spec: m: 128; k: 1536; l: 1; seed: 1111
test.1.status: pass
test.2.spec: m: 128; k: 3072; l: 1; seed: 1111
test.2.status: pass
test.3.spec: m: 256; k: 7168; l: 1; seed: 1111
test.3.status: pass
test.4.spec: m: 256; k: 7168; l: 1; seed: 1111
test.4.status: pass
test.5.spec: m: 2432; k: 4608; l: 2; seed: 1111
test.5.status: pass
test.6.spec: m: 384; k: 7168; l: 2; seed: 1111
test.6.status: pass
test.7.spec: m: 512; k: 512; l: 2; seed: 1111
test.7.status: pass
test.8.spec: m: 512; k: 4096; l: 2; seed: 1111
test.8.status: pass
test.9.spec: m: 512; k: 1536; l: 2; seed: 1111
test.9.status: pass
check: pass

For NVFP4 GEMM (using tensor-core)

reference-kernels/problems/nvidia/nvfp4_gemm> python3 eval.py test task.yml
compile: start
compile: pass
test-count: 10
test.0.status: pass
test.1.spec: m: 128; n: 1536; k: 7168; l: 1; seed: 1111
test.1.status: pass
test.2.spec: m: 128; n: 3072; k: 1536; l: 1; seed: 1111
test.2.status: pass
test.3.spec: m: 256; n: 7168; k: 256; l: 1; seed: 1111
test.3.status: pass
test.4.spec: m: 256; n: 7168; k: 2048; l: 1; seed: 1111
test.4.status: pass
test.5.spec: m: 2304; n: 4608; k: 7168; l: 1; seed: 1111
test.5.status: pass
test.6.spec: m: 384; n: 7168; k: 2304; l: 1; seed: 1111
test.6.status: pass
test.7.spec: m: 512; n: 512; k: 7168; l: 1; seed: 1111
test.7.status: pass
test.8.spec: m: 512; n: 4096; k: 512; l: 1; seed: 1111
test.8.status: pass
test.9.spec: m: 512; n: 1536; k: 7168; l: 1; seed: 1111
test.9.status: pass
check: pass

For NVFP4 dual_gemm(using tensor-core)

reference-kernels/problems/nvidia/nvfp4_dual_gemm>  python3 eval.py test task.yml
compile: start
compile: pass
test-count: 10
test.0.spec: m: 128; n: 256; k: 256; l: 1; seed: 1111
test.0.status: pass
test.1.spec: m: 128; n: 1536; k: 7168; l: 1; seed: 1111
test.1.status: pass
test.2.spec: m: 128; n: 3072; k: 1536; l: 1; seed: 1111
test.2.status: pass
test.3.spec: m: 256; n: 7168; k: 256; l: 1; seed: 1111
test.3.status: pass
test.4.spec: m: 256; n: 7168; k: 2048; l: 1; seed: 1111
test.4.status: pass
test.5.spec: m: 2304; n: 4608; k: 7168; l: 1; seed: 1111
test.5.status: pass
test.6.spec: m: 384; n: 7168; k: 2304; l: 1; seed: 1111
test.6.status: pass
test.7.spec: m: 512; n: 512; k: 7168; l: 1; seed: 1111
test.7.status: pass
test.8.spec: m: 512; n: 4096; k: 512; l: 1; seed: 1111
test.8.status: pass
test.9.spec: m: 512; n: 1536; k: 7168; l: 1; seed: 1111
test.9.status: pass
check: pass

For NVFP4 group gemm(using tensor-core)

reference-kernels/problems/nvidia/nvfp4_group_gemm>
compile: start
compile: pass
test-count: 10
test.0.spec: m: 128; n: 256; k: 512; g: 8; seed: 1111
test.0.status: pass
test.1.spec: m: 128; n: 256; k: 512; g: 2; seed: 1111
test.1.status: pass
test.2.spec: m: 128; n: 384; k: 640; g: 3; seed: 1111
test.2.status: pass
test.3.spec: m: 256; n: 384; k: 640; g: 4; seed: 1111
test.3.status: pass
test.4.spec: m: 256; n: 512; k: 384; g: 2; seed: 1111
test.4.status: pass
test.5.spec: m: 384; n: 512; k: 384; g: 2; seed: 1111
test.5.status: pass
test.6.spec: m: 384; n: 640; k: 512; g: 2; seed: 1111
test.6.status: pass
test.7.spec: m: 256; n: 640; k: 128; g: 8; seed: 1111
test.7.status: pass
test.8.spec: m: 512; n: 768; k: 256; g: 5; seed: 1111
test.8.status: pass
test.9.spec: m: 512; n: 768; k: 768; g: 3; seed: 1111
test.9.status: pass
check: pass