[BugFix] Use x*x*x instead of pow(x,3)

[CharlieFRuan]

This PR changes topi.power(x,3) in gelu_tanh to x * x * x. Subsequently, we add a warning when pow(x, y), where y >= 3, is detected.

This is motivated by, in ROCm, pow(x, y) returns NaN when x < 0 and y >= 3. This is shown below:

@I.ir_module
class Module:
    @T.prim_func
    def pow3(A: T.Buffer((1,), "float16")):
        T.func_attr({"tir.is_scheduled": 1, "tir.noalias": T.bool(True)})
        # with T.block("root"):
        for ax0_fused_0 in T.thread_binding(1, thread="blockIdx.x"):
            for ax0_fused_1 in T.thread_binding(1024, thread="threadIdx.x"):
                with T.block("T_multiply_3"):
                    T.where(ax0_fused_0 * 1024 + ax0_fused_1 < 1)
                    T.reads(A[0])
                    T.writes(A[0])
                    # A[0] = T.pow(A[0], T.float16(2))  # Gives 1
                    # A[0] = T.pow(T.float16(-1), T.float16(3))  # Gives -1
                    # A[0] = T.pow(A[0], T.int32(3))  # gives nan
                    A[0] = T.pow(A[0], T.float16(3))  # gives nan
                    # A[0] = T.pow(A[0], T.float16(4.0))  # gives nan

mod = Module
rt_mod = tvm.build(mod, target="rocm")
A_tvm = tvm.nd.array(np.array([-1.0], dtype="float16"), tvm.rocm())
print(f"Before: {A_tvm.numpy()}")
rt_mod["pow3"](A_tvm)
print(f"After: {A_tvm.numpy()}")

[CharlieFRuan]

cc @tqchen @MasterJH5574 @junrushao