extra/gemm/max_matmul: start of custom kernels for GEMM

[flammit]

As a prerequisite to implementing full speed GEMM for NV, here's the handwritten versions of GEMM that shows the incremental progress needed to get there and associated speed improvements.

Acc	Variation	Performance
FP32	hcopt	1301.50 us, would be 105600.10 GFLOPS matmul, 77.34 GB/s
FP32	flat_smem_input	1309.41 us, would be 104962.67 GFLOPS matmul, 76.88 GB/s
FP32	swizzled_smem_input	882.69 us, would be 155705.02 GFLOPS matmul, 114.04 GB/s
FP32	2_stage_swizzled_smem_input	831.49 us, would be 165292.77 GFLOPS matmul, 121.06 GB/s
FP32	max	826.37 us, would be 166316.89 GFLOPS matmul, 121.81 GB/s
FP16	3_stage_swizzled	505.66 us, would be 271798.97 GFLOPS matmul, 199.07 GB/s
FP16	max	404.48 us, would be 339791.71 GFLOPS matmul, 248.87 GB/s

The command for hcopt is: PYTHONPATH=. CUDA=1 GEMM_VARIATION="hcopt" DTYPE_IN=half DTYPE_OUT=half DTYPE_ACC=float CNT=1024 INPUT=RAND python3 ./extra/gemm/max_matmul.py .

The command for the rest of the FP32 acc is: PYTHONPATH=. CUDA=1 GEMM_VARIATION="$VARIATION" DTYPE_IN=half DTYPE_OUT=float DTYPE_ACC=float CNT=1024 INPUT=RAND python3 ./extra/gemm/max_matmul.py.

The command for the rest of the FP16 acc is: PYTHONPATH=. CUDA=1 GEMM_VARIATION="$VARIATION" DTYPE_IN=half DTYPE_OUT=half DTYPE_ACC=half CNT=1024 INPUT=ONES python3 ./extra/gemm/max_matmul.py.

[flammit]

added an unoptimized FP16 input/FP16 acc MMA example

[geohot]

So I tested this with Triton once, what does Triton get? IIRC it was well over 200. We should at least match that with handcoded stuff.

[github-actions]

This branch currently is behind tinygrad/master. The line count difference bot is disabled.

[flammit]

Added a 3-staged pipeline with swizzled SMEM inputs example for FP16 acc that does 270TF (still less than the 330TF in cutlass, but not as bad as before).

note this PR depends on #6956 being landed first.

[flammit]

happy to remove these extra stages/variations and just keep the "max" variations, but figure it might be useful for comparison as future kernel rendering features are incrementally added.