[CuTe] adding DSL INT8 MMA support on SM80+, with CuTe C++ example

[ayghri]

This PR adds INT8 warp-level MMA with an Ampere batched GEMM example. Partial solution to #3081

• Add MmaI8Op and MmaIntOverflow in python/CuTeDSL/cutlass/cute/nvgpu/warp/mma.py, wraps mma.sync.aligned with .s8/.u8 qualifiers, shapes (16,8,16) and
  (16,8,32), saturate/wrap overflow modes.
• examples/python/CuTeDSL/ampere/tensorop_gemm_i8.py. Follows tensorop_gemm.py conventions.
• examples/cute/tutorial/igemm_sm80.cu. CuTE C++ implementation, follows sgemm_sm80.cu.

What works:

• S8xS8 signedness, --b_dtype U8 kept for future IR support
• Arbitrary M, N, K (tensors padded to tile boundaries)
• N-major or M-major output C
• Auto-selected BM tile (16/32/64/128) based on M dimension

What does not work:

• DSL: S8xU8 supported by hardware but blocked by DSL compiler.

INT8 vs FP16 differences:

• K-major inputs only: ldmatrix requires 128-bit aligned smem. FP16
  column-major gives 8×16b = 128 bits; INT8 column-major gives 8×8b = 64 bits.
• Swizzle<2,4,3> instead of Swizzle<bits,3,3>: INT8 needs MBase=4
  (not 3) because log2(128/8) = 4.
• Direct gmem epilogue: INT32 output makes the smem C buffer 128×128×4B =
  64KB, limiting occupancy to 1 CTA/SM (for my RTX 3090, I haven't benchmarked on A100 yet).
• Python example script auto-selects BM tile based on my local benchmarks

Performance vs PyTorch 2.9.1+cu128 (torch._int_mm), RTX 3090

100 iterations, CUDA event timing, TOPS = 2MN*K/time.

N = K = 4096

M	PyTorch (TOPS)	CuTe DSL (TOPS)	vs PyTorch
16	n/a	18.8	n/a
32	7.0	35.9	514%
64	13.8	60.1	435%
128	27.5	84.5	307%
256	55.0	165.0	300%
512	56.5	179.7	318%
1024	57.1	183.6	322%
2048	65.6	218.9	334%
4096	74.1	226.7	306%

N = K = 16384

M	PyTorch (TOPS)	CuTe DSL (TOPS)	vs PyTorch
32	15.4	49.5	321%
64	28.9	101.2	350%
128	57.6	168.9	293%
256	58.3	179.8	309%
512	69.3	212.7	307%
1024	73.1	225.9	309%
4096	75.0	230.2	307%

Non-aligned dimensions

M	N	K	PyTorch (TOPS)	CuTe DSL (TOPS)	vs PyTorch
33	4096	4096	7.4	32.7	440%
65	4096	4096	15.1	46.0	305%
200	4096	4096	46.2	138.2	299%
128	1000	4096	7.3	21.1	289%
128	4096	1000	24.2	65.4	270%

benchmarks show cuBLAS within ~1-25% of this kernel for most sizes.

Testing

python examples/python/CuTeDSL/ampere/tensorop_gemm_i8.py --mnkl 512,512,512,1
python examples/python/CuTeDSL/ampere/tensorop_gemm_i8.py --mnkl 33,200,100,1
python examples/python/CuTeDSL/ampere/tensorop_gemm_i8.py --mnkl 4096,4096,4096,1 --skip_ref_check --iterations 100