[bug] Use GPUDEBUG macro instead of raw printf for kernel error message

[PrasoonMishra]

The raw printf in the "should never happen" branch is compiled into the kernel even in release builds, increasing register pressure. Since MCX is register-limited, replacing it with GPUDEBUG (a no-op unless MCX_DEBUG is defined) frees registers and improves occupancy, giving ~2.5% speedup in release mode on AMD gf90a.

[fangq]

use printf for printing "should never happen" is by design. this is a critical message that must be printed, if ever happens, regardless if the binary was compiled in debug mode or not.

[PrasoonMishra]

@fangq
Thanks, that makes sense.

If above design is by choice, then I would like to flag for consistency issue between this and OpenCL port. The OpenCL port handles this same check differently. There it's gated behind GPUDEBUG, so in a normal build it's compiled out and never prints. ( mcx_core.cl Line: 3289 )

if ((GPU_PARAM(gcfg, debuglevel) & MCX_DEBUG_MOVE_ONLY) &
        (mediaid == 0 || idx1d == OUTSIDE_VOLUME_MIN || idx1d == OUTSIDE_VOLUME_MAX)) {
    GPUDEBUG(("ERROR: should never happen! mediaid=%d idx1d=%X\n", mediaid, idx1d));
    return;
}

As a result, the two ports currently diverge in behavior for the same condition. It would be great if they could be aligned so both ports handle this case consistently.