babl: decrease tolerance

babl: decrease tolerance

The following conversion is undesirable, going via linear u16 on the
way to a different TRC u16 amount of quantization should be considered
bad enough to be avoided.

chosen RGBA u32 to R'G'B'A u16: steps: 4 error: 0.000005 cost: 876.000000
	/usr/local/lib/babl-0.1/u32.so 0: RGBA u32 to RGBA u16
	/usr/local/lib/babl-0.1/gggl-table-lies.so 0: RGBA u16 to RGBA float
	/usr/local/lib/babl-0.1/sse2-float.so 0: RGBA float to R'G'B'A float
	/usr/local/lib/babl-0.1/gggl.so 0: R'G'B'A float to R'G'B'A u16

This is slower but more correct:

chosen RGBA u32 to R'G'B'A u16: steps: 3 error: 0.000000 cost: 1051.000000
	/usr/local/lib/babl-0.1/simple.so 0: RGBA u32 to RGBA float
	/usr/local/lib/babl-0.1/sse2-float.so 0: RGBA float to R'G'B'A float
	/usr/local/lib/babl-0.1/gggl.so 0: R'G'B'A float to R'G'B'A u16