rms_norm example doesn't support float16

[ChaoyangJi]

for (int32_t by_1 = 0; by_1 < 4: ++by_1) {
tl::ascend:copy_gm_to_ub<half, 64, 32>(a_ub[0], A[((cid * 16384) + (vid * 8192)) + (by_1 * 64))], 256);
AscendC:PipeBarrier():
for (int32 t i = 0; i < 32; ++i) {
AscendC::PipeBarriersPIPE AL>0:
auto mean_square_ub_scalar = 1.0f / mean_square_ub.GetValue(i);
Ascendc:Muls(a ub[(1 * 64)], a_ub[(i( * 64)], mean_square_ub_scalar, 64);
AscendC:PipeBarrier<PIPE_ALL>();
}
tl::ascend:copy_ub_to_gm<half, 64, 32>(B[((cid * 16384) + (vid * 8192)) + (by_1 * 64))], a_ub[0], 256);
AscendC:PipeBarrier<PIPE_ALL>() ;
}

auto mean_square_ub_scalar = 1.0f / mean_square_ub.GetValue(i);

/tmp/tmpnqOuzobk.cpp:64:43: error:half precision operation is not allowed in aicore function
Fatal Err: A fatal error occurred in iternel process, and Diagnostic info can not be reported correctly

[liggest]

According to yesterday's discussion, if mean_square_ub.GetValue(i) returns half, maybe it should be cast to float before division by 1.0f. Then the result auto mean_square_ub_scalar should be cast back to half.