ac/gpu_info: Use most significant SE mask bit for num_se
Using the bit count is flawed when there are SEs with a 0 CU mask.
The CU mask for a 7900GRE looks like this:
cu_mask[SE0][SA0] = 0xff
cu_mask[SE0][SA1] = 0xff
cu_mask[SE1][SA0] = 0xff
cu_mask[SE1][SA1] = 0xff
cu_mask[SE2][SA0] = 0xff
cu_mask[SE2][SA1] = 0xff
cu_mask[SE3][SA0] = 0xff
cu_mask[SE3][SA1] = 0xff
cu_mask[SE4][SA0] = 0x0
cu_mask[SE4][SA1] = 0x0
cu_mask[SE5][SA0] = 0xff
cu_mask[SE5][SA1] = 0xff
We should treat this as 6 SEs for the purpose of scratch allocation etc., otherwise there can be hangs.