Fix bogus span mask in s_atifragshader.c leading to trouble when the pixel was later discarded by z test (this fixes doom3 r200 renderpath with swrast). Fix calling _swrast_alpha_test potentially leading to bogus results when Alpha Test might not be enabled. While here, don't disable deferred_texture (early z) when ATI_fragment_shader is active, as it is not necessary (for the record, this boosts doom3 swrast performance by roughly a factor of 2 (4 seconds per frame now...) with the r200 render path, might be a worthwile optimization for fragment programs which don't write z).