[XLA:GPU] Fix Cublaslt fp8 to dummy nullptr c pointer when beta = 0

[Cjkkkk]

📝 Summary of Changes
Fix Cublaslt fp8 to dummy nullptr c pointer when beta = 0

🎯 Justification
Since cuda 13.0, cublasLT adds a check that enforce gemm with same C and D pointer to have same descs. XLA uses same C and D pointer for in-place gemms which is fine. But for gemms with beta = 0, XLA also uses same C and D pointer even though C pointer is not needed by cublas. The new check will fail now in this case for fp8 gemms since C and D usually have different descs. The correct way is to use dummy C nullptr for gemms with beta = 0.

🚀 Kind of Contribution
🐛 Bug Fix

📊 Benchmark (for Performance Improvements)
None

🧪 Unit Tests:
these tests will pass with cuda 13 now:

DotTests/ParametricDotTest.TestF8E4M3FN/270x270x520_MajorToMinorFT
DotTests/ParametricDotTest.TestF8E4M3FN/12x117x7_MajorToMinorFT
DotTests/ParametricDotTest.TestF8E4M3FN/260x3x520_MajorToMinorTT
DotTests/ParametricDotTest.TestF8E4M3FN/12x117x7_MajorToMinorTT
DotTests/ParametricDotTest.TestF8E4M3FN/260x3x520_MajorToMinorTF
DotTests/ParametricDotTest.TestF8E4M3FN/260x3x520_MajorToMinorFT
DotTests/ParametricDotTest.TestF8E4M3FN/12x117x7_MajorToMinorTF
DotTests/ParametricDotTest.TestF8E4M3FN/12x117x7_MajorToMinorFF
DotTests/ParametricDotTest.TestF8E4M3FN/270x270x520_MajorToMinorTF
DotTests/ParametricDotTest.TestF8E4M3FN/260x3x520_MajorToMinorFF
DotTests/ParametricDotTest.TestF8E4M3FN/270x270x520_MajorToMinorFF
DotTests/ParametricDotTest.TestF8E4M3FN/270x270x520_MajorToMinorTT
F8E5M2Tests/DotAlgorithmSupportTest.AlgorithmIsSupportedFromCudaCapability/dot_any_f8_any_f8_f32_with_lhs_f8e5m2_rhs_f8e4m3fn_output_f8e4m3fn_from_cc_8_9_rocm_63_no_restriction_c_16_nc_2
F8E4M3FNTests/DotAlgorithmSupportTest.AlgorithmIsSupportedFromCudaCapability/dot_any_f8_any_f8_f32_fast_accum_with_lhs_f8e4m3fn_rhs_f8e4m3fn_output_f8e4m3fn_from_cc_8_9_rocm_63_no_restriction_c_16_nc_2
F8E5M2Tests/DotAlgorithmSupportTest.AlgorithmIsSupportedFromCudaCapability/dot_any_f8_any_f8_f32_with_lhs_f8e5m2_rhs_f8e4m3fn_output_f8e4m3fn_from_cc_8_9_rocm_63_no_restriction_c_32_nc_32
F8E4M3FNTests/DotAlgorithmSupportTest.AlgorithmIsSupportedFromCudaCapability/dot_any_f8_any_f8_f32_with_lhs_f8e4m3fn_rhs_f8e4m3fn_output_f8e4m3fn_from_cc_8_9_rocm_63_no_restriction_c_32_nc_32
F8E5M2Tests/DotAlgorithmSupportTest.AlgorithmIsSupportedFromCudaCapability/dot_any_f8_any_f8_f32_fast_accum_with_lhs_f8e5m2_rhs_f8e4m3fn_output_f8e4m3fn_from_cc_8_9_rocm_63_no_restriction_c_16_nc_2
F8E5M2Tests/DotAlgorithmSupportTest.AlgorithmIsSupportedFromCudaCapability/dot_any_f8_any_f8_f32_with_lhs_f8e5m2_rhs_f8e4m3fn_output_f8e5m2_from_cc_8_9_rocm_63_no_restriction_c_16_nc_2
F8E4M3FNTests/DotAlgorithmSupportTest.AlgorithmIsSupportedFromCudaCapability/dot_any_f8_any_f8_f32_with_lhs_f8e4m3fn_rhs_f8e4m3fn_output_f8e4m3fn_from_cc_8_9_rocm_63_no_restriction_c_16_nc_2
F8E4M3FNTests/DotAlgorithmSupportTest.AlgorithmIsSupportedFromCudaCapability/dot_any_f8_any_f8_f32_fast_accum_with_lhs_f8e4m3fn_rhs_f8e4m3fn_output_f8e4m3fn_from_cc_8_9_rocm_63_no_restriction_c_32_nc_32
F8E5M2Tests/DotAlgorithmSupportTest.AlgorithmIsSupportedFromCudaCapability/dot_any_f8_any_f8_f32_with_lhs_f8e5m2_rhs_f8e4m3fn_output_f8e5m2_from_cc_8_9_rocm_63_no_restriction_c_32_nc_32
F8E5M2Tests/DotAlgorithmSupportTest.AlgorithmIsSupportedFromCudaCapability/dot_any_f8_any_f8_f32_fast_accum_with_lhs_f8e5m2_rhs_f8e4m3fn_output_f8e5m2_from_cc_8_9_rocm_63_no_restriction_c_16_nc_2
F8E5M2Tests/DotAlgorithmSupportTest.AlgorithmIsSupportedFromCudaCapability/dot_any_f8_any_f8_f32_fast_accum_with_lhs_f8e5m2_rhs_f8e4m3fn_output_f8e5m2_from_cc_8_9_rocm_63_no_restriction_c_32_nc_32
F8E5M2Tests/DotAlgorithmSupportTest.AlgorithmIsSupportedFromCudaCapability/dot_any_f8_any_f8_f32_fast_accum_with_lhs_f8e5m2_rhs_f8e4m3fn_output_f8e4m3fn_from_cc_8_9_rocm_63_no_restriction_c_32_nc_32

🧪 Execution Tests:
None

[Cjkkkk]

Hi @beckerhe , could you help take a look at this? This impacts lots of fp8 tests on our CUDA 13.1 release.