Fix crash on non-empty return_void blocks

[ShangkunLi]

The --canonicalize-return pass previously assumed that every block containing a neura.return_void operation was an empty block (i.e., return_void was the only operation in it). When this assumption was violated — which happens for real-world kernels where other operations share the same basic block as return_void — the pass hit an unconditional assert(false) and crashed:

assert(false && "Unsupported case: return block is not empty.");

This made the pass fail on a broad class of benchmarks (fir, conv, spmv, jacobi, etc.) that produce non-empty return_void blocks after lowering.

Fix:

• Removed the is_empty_block guard and the failing assert in both call sites inside processVoidReturnsInKernel and CanonicalizeReturnPass.
• Renamed processEmptyReturnVoidBlock → processReturnVoidBlock to reflect that it now handles the general case.
• The underlying predecessor-walking logic already handled non-empty blocks correctly; the guard was overly conservative.

[tancheng]

Why we don't need to update any tests? They all already passed?

[tancheng]

Why we previously want a standalone bb only containing return? Can you recall?