shaders/sampling: don't use polar compute for high radius

After a certain input size, the gather kernel pulls ahead. This is
either due to interactions with the size of the texture cache, or due to
alignment with the 2x2 gather input grid, or both.

Regardless, this should generally speed things up for very high radius
polar kernels (especially when downscaling).

src/shaders/sampling.c

@@ -664,7 +664,11 @@ bool pl_shader_sample_polar(pl_shader sh, const struct pl_sample_src *src,
     int bw = 32, bh = sh_glsl(sh).max_group_threads / bw;
     int sizew, sizeh, iw, ih;
 
+    // Disable compute shaders after a (hard-coded) radius of 6, since the
+    // gather kernel generally pulls ahead here.
     bool is_compute = !params->no_compute && sh_glsl(sh).compute;
+    is_compute &= obj->filter->radius < 6.0;
+
     while (is_compute) {
         // We need to sample everything from base_min to base_max, so make sure
         // we have enough room in shmem. The extra margin on the ceilf guards