[BugFix][TIRx] Fix bad-optional-access in BF16/FP8 legalize passes for target-less PrimFuncs

[swjng]

Problem

BF16ComputeLegalize, BF16StorageLegalize, FP8ComputeLegalize, and FP8StorageLegalize all call f->GetAttr<Target>(kTarget).value() unconditionally. Host-side helper PrimFuncs produced during Relax lowering (e.g. for reshape, mean) carry no target attribute; .value() on an empty Optional<Target> aborts at runtime:

terminate called after throwing an instance of 'std::bad_optional_access'

This surfaces whenever tvm.compile targets CUDA on a model with BF16/FP8 support compiled in, because the Relax lowering pipeline mixes target-annotated GPU kernels with target-less CPU helpers in the same module.

Fix

Retrieve the target into opt_target and combine the defined() check with CheckDataTypeSupport:

if (opt_target.defined() && CheckDataTypeSupport(opt_target.value(), "tvm.contrib.nvcc.supports_bf16")) {
    return f;
}

• If the target is absent: legalization runs. This is safe — it is a no-op when no BF16/FP8 types are present, and it ensures correctness for the edge case where a target-less function does contain such types.
• If the target confirms native support: legalization is skipped as before.
• If the target lacks native support: legalization runs as before.

This is consistent with how explicit host targets (e.g. llvm) are handled, which also go through legalization. Applied identically to all four legalize pass lambdas.

[gemini-code-assist]

Code Review

This pull request updates the BF16 and FP8 legalization passes to safely handle functions where the target attribute is missing, preventing potential crashes. However, the current implementation returns early when a target is undefined, which inadvertently skips legalization for target-less functions. Feedback suggests combining the target existence check with the capability check to ensure consistency with host targets and to allow the legalizer to run on functions without an explicit target.