Fix -mfloat-abi=soft compilation for ARM with OpenCL target

[trevor-m]

Currently, any extra attributes of the llvm target are not being passed through to LLVMModule during ModulePackImportsToLLVM. This will cause compilation to fail with error: /tmp/tmpctakbpk1/devc.o uses VFP register arguments, output does not when the target is opencl and the target_host requires -mfloat-abi=soft. This regression was introduced by #4657.

There are two reasons why:

1. python/tvm/runtime/module.py uses "_get_target_triple" to get the target triple to use for ModulePackImportsToLLVM. However, the implementation of that function returns tm_->getTargetTriple().str() which will only have the standard <arch><sub>-<vendor>-<sys>-<abi> and is missing any extra flags or attributes such as -mfloat-abi=soft.
2. Since the module was created by CodegenBlob, it doesn't store anything in the "tvm_target" module metadata and therefore module_->getTargetTriple() is called which again doesn't have the extra flags and attributes: https://github.com/apache/incubator-tvm/blob/master/src/target/llvm/llvm_module.cc#L254

This PR fixes those problems by:

1. Adding important attributes back into _get_target_triple (only mfloat-abi for now). This part doesn't seem ideal to me. Any thoughts on how to improve? Can we store the user's full target string earlier on?
2. Saving full target triple string in "tvm_target" module metadata flag during CodeGenBlob so that LLVMModuleNode::Init(std::unique_ptr<llvm::Module> module, std::shared_ptr<llvm::LLVMContext> ctx) can retrieve it: https://github.com/apache/incubator-tvm/blob/master/src/target/llvm/llvm_module.cc#L247 This matches what LLVMModuleNode::Init(const IRModule& mod, std::string target) already does.

Discuss post: https://discuss.tvm.ai/t/opencl-target-for-32-bit-arm-linux-android-broken-after-pr-4657/7252
Fixes #6019

[junrushao]

LGTM. Thanks Trevor!

[kevinthesun]

LGTM

[trevor-m]

Looks like some unrelated failures in ci? tests/python/frontend/tensorflow/test_forward.py::test_forward_nms FAILED [ 50%]

E     let %x7: Tensor[(5), int32] = vm.reshape_tensor(%x5, meta[relay.Constant][6], meta[relay.attrs.ReshapeTensorAttrs][0]) an internal invariant was violated while typechecking your program [00:42:28] /workspace/src/relay/op/vm/vm.cc:195: Check failed: tt: input must be tensor type

[junrushao]

Yes, looks like the CI failure is not related

[trevor-m]

@kevinthesun CI is passing now