Fix GPU detection in PerStoreFeatureNode

[moonquekes]

Bug Description: Incorrect Assignment of is_gpu Property

Summary:
In the context of feature extraction using meta_schedule, the is_gpu property was incorrectly assigned by directly comparing the target kind name to "cuda". This approach is not only too specific but also potentially unsafe, as it does not account for different GPU types or handle cases where the target might be undefined.

Incorrect Code:

tvm/src/meta_schedule/feature_extractor/per_store_feature.cc

Line 1395 in 567eeed

bool is_gpu = tune_context->target.value()->kind->name == "cuda";

bool is_gpu = tune_context->target.value()->kind->name == "cuda";

Problems with the Incorrect Code:

1. Lack of Generality: Not all GPUs use CUDA (e.g., AMD GPUs may use ROCm or OpenCL).
2. Specificity Issue: Even on NVIDIA hardware, the target kind name isn't guaranteed to be "cuda" and can vary based on library versions or configurations.

Correct Implementation:

auto& target_keys = tune_context->target.value()->keys;
bool is_gpu = std::find(target_keys.begin(), target_keys.end(), "gpu") != target_keys.end();

Explanation of Correct Code:
The corrected code checks if "gpu" exists within the list of keys associated with the target. This method is more robust because:

• It correctly identifies any target that includes "gpu" in its keys, regardless of the underlying hardware or software platform.
• It avoids potential null pointer dereferencing by not directly accessing kind->name.
• It is more flexible and can accommodate various types of GPU targets without hardcoding specific names.

[moonquekes]

hi，@Hzfengsy.
Please review the code I submitted, thanks!

[Hzfengsy]

Thanks for the fix

[yongwww]

@tvm-bot rerun