#3079 prefill: check both f16-range symbol writes on the reset path

Descrição

One change in `src/prefill/kernels.cu` (5 lines): `f16_range_read(reset=true)` fired the two `cudaMemcpyToSymbol` writes and never looked at the result. When one failed, the symbol kept its previous value and the next read reported a stale cumulative count with nothing saying the reset did not happen.

Now both writes always run (a failed first write must not skip the second), both results are checked, and failure is reported.

**What it leaves alone**: behaviour is unchanged when both succeed — the normal path still does exactly two symbol writes and one error check, and the reads are untouched. No defaults change.

Not tested: a CUDA build (no CUDA toolkit on this box); the change compiles on the gfx906 HIP target.