fix: fix cublas_scaled_mm

[dc3671]

In pytorch 2.7, cublas's workspace is modified from 32MB to 1MB on Hopper:

• Use _getWorkspaceSize() rather than _getWorkspace() pytorch/pytorch@203a27e#diff-74fcb26047c1df4024105d36ce22a36b77cf8cc93c28631d743e639b3d6066aeL1624-R1599

-   auto stream = c10::cuda::getCurrentCUDAStream();
-   size_t workspaceSize = 0;
-   auto workspace_ptr = _getWorkspace(workspaceSize);
+   size_t workspaceSize = _getWorkspaceSize();
+   auto workspace = at::empty(static_cast<int64_t>(workspaceSize),    at::TensorOptions().dtype(at::kByte).device(at::kCUDA));

• And in _getWorkspace() will use cublas's workspace size 32MB on Hopper: https://github.com/pytorch/pytorch/blob/main/aten/src/ATen/cuda/CublasHandlePool.cpp#L133

const size_t default_size = sm90 ? 4096 * 8 * 1024 : 4096 * 1024 * 2 + 16 * 1024 * 8;

And this lead to different algos (especially when one choose splitK and another not) chosen by cublasLt on some GEMM shapes. So we need to align this in our tests.

In E2E test, there are other tests called torch._scaled_mm, and the workspace size variable is static in cpp code, so the modification in test_scaled_mm won't take effect and we need to add that in test_linear_fp8.py.

[dc3671]

/bot run --stage-list "H100_PCIe-PyTorch-1"

[tensorrt-cicd]

PR_Github #2412 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #2412 [ run ] completed with state FAILURE
/LLM/main/L0_MergeRequest_PR pipeline #1733 (Partly Tested) completed with status: 'FAILURE'

[dc3671]

/bot run --stage-list "H100_PCIe-PyTorch-1"

[tensorrt-cicd]

PR_Github #2420 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #2420 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #1738 (Partly Tested) completed with status: 'FAILURE'

[dc3671]

/bot run --stage-list "H100_PCIe-PyTorch-1"

[tensorrt-cicd]

PR_Github #2430 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #2430 [ run ] completed with state FAILURE
/LLM/main/L0_MergeRequest_PR pipeline #1746 (Partly Tested) completed with status: 'FAILURE'

[dc3671]

/bot run --disable-fail-fast --stage-list "H100_PCIe-PyTorch-1"

[tensorrt-cicd]

PR_Github #2451 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #3 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #3 [ run ] completed with state ABORTED

[tensorrt-cicd]

PR_Github #2451 [ run ] completed with state FAILURE

[yiqingy0]

/bot run --disable-fail-fast --stage-list "H100_PCIe-PyTorch-1"

[tensorrt-cicd]

PR_Github #2453 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #6 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #6 [ run ] completed with state ABORTED

[dc3671]

/bot run --disable-fail-fast --stage-list "H100_PCIe-PyTorch-1"

[tensorrt-cicd]

PR_Github #2459 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #2453 [ run ] completed with state ABORTED

[tensorrt-cicd]

PR_Github #2459 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #1763 (Partly Tested) completed with status: 'FAILURE'

[dc3671]

/bot run --disable-fail-fast --stage-list "H100_PCIe-PyTorch-1"

[tensorrt-cicd]

PR_Github #2595 [ run ] triggered by Bot

[dc3671]

/bot run --disable-fail-fast --stage-list "H100_PCIe-PyTorch-1"

[tensorrt-cicd]

PR_Github #2595 [ run ] completed with state FAILURE
/LLM/main/L0_MergeRequest_PR pipeline #1852 (Partly Tested) completed with status: 'FAILURE'

[tensorrt-cicd]

PR_Github #2615 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #2615 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #1862 (Partly Tested) completed with status: 'SUCCESS'

[dc3671]

/bot run --disable-fail-fast --stage-list "H100_PCIe-PyTorch-1"

[tensorrt-cicd]

PR_Github #2703 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #2703 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #1928 (Partly Tested) completed with status: 'FAILURE'

[dc3671]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #2821 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #2821 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #1992 completed with status: 'SUCCESS'

[dc3671]

/bot reuse-pipeline

[tensorrt-cicd]

PR_Github #2910 [ reuse-pipeline ] triggered by Bot

[tensorrt-cicd]

PR_Github #2910 [ reuse-pipeline ] completed with state SUCCESS
Reusing PR_Github #2821 for commit 35452ee