GPU: Handle bad fog params as large signed vals

[unknownbrackets]

From tests, it seems they're just treated as valid exponents.

Using 65535 since that's the range of depth, can't think of a game using a larger value for a fog parameter, probably never even this large.

Did some more fog tests, I think I had a mistake before when I thought softgpu fog was reversed - but since I was there, I validated how inf/nan behave. It seems like they act just like vfpu min/max - the exponent is just not treated as special.

This could use validation against MediEvil and any other affected games, though. But this matches tests.

-[Unknown]

[hrydgard]

This is probably just as fine as the previous value of 100000 for coef 0, but it's conceivable that games would not want the fog to fully fade in at the end of depth. That would be pretty strange usage though, so.... let's try it.

[unknownbrackets]

Right. I'm not opposed to a different constant - was just extending it to handle -NAN and INF for both registers. I actually didn't notice that 100000 wasn't just 10000 used in the INF branch of the prev code (note: both were only for coef 1.)

-[Unknown]