Partly SIMDify the `f32_to_u8` method (#1102)

This gives me -7% runtime on the bilinear image rendering benchmark on
NEON:

```
     Running benches/main.rs (/Users/lstampfl/Programming/GitHub/vello/target/release/deps/main-011ad84cd531a979)
fine/image/quality/medium_u8_neon
                        time:   [3.4155 µs 3.4190 µs 3.4231 µs]
                        change: [-7.3623% -7.1971% -7.0174%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 7 outliers among 100 measurements (7.00%)
  2 (2.00%) high mild
  5 (5.00%) high severe
```

On WASM, it drops the frame time from 40ms to 30ms, so a huge
difference! I'm not sure if it's possible to easily SIMDify the u32 to
u8 conversion (at least from what I can tell `tiny-skia` also only uses
SIMD for the f32 to u32 conversion), open to suggestions, though.

sparse_strips/vello_cpu/src/fine/mod.rs

diff --git a/sparse_strips/vello_cpu/src/fine/mod.rs b/sparse_strips/vello_cpu/src/fine/mod.rs
index 07f849b..9feb999 100644
--- a/sparse_strips/vello_cpu/src/fine/mod.rs
+++ b/sparse_strips/vello_cpu/src/fine/mod.rs
@@ -91,24 +91,29 @@
 
 #[inline(always)]
 pub(crate) fn f32_to_u8<S: Simd>(val: f32x16<S>) -> u8x16<S> {
-    // TODO: SIMDify
+    let simd = val.simd;
+    // Note that converting to u32 first using SIMD and then u8
+    // is much faster than converting directly from f32 to u8.
+    let converted = simd.cvt_u32_f32x16(val);
+
+    // TODO: Maybe we can also do this using SIMD?
     [
-        val.val[0] as u8,
-        val.val[1] as u8,
-        val.val[2] as u8,
-        val.val[3] as u8,
-        val.val[4] as u8,
-        val.val[5] as u8,
-        val.val[6] as u8,
-        val.val[7] as u8,
-        val.val[8] as u8,
-        val.val[9] as u8,
-        val.val[10] as u8,
-        val.val[11] as u8,
-        val.val[12] as u8,
-        val.val[13] as u8,
-        val.val[14] as u8,
-        val.val[15] as u8,
+        converted[0] as u8,
+        converted[1] as u8,
+        converted[2] as u8,
+        converted[3] as u8,
+        converted[4] as u8,
+        converted[5] as u8,
+        converted[6] as u8,
+        converted[7] as u8,
+        converted[8] as u8,
+        converted[9] as u8,
+        converted[10] as u8,
+        converted[11] as u8,
+        converted[12] as u8,
+        converted[13] as u8,
+        converted[14] as u8,
+        converted[15] as u8,
     ]
     .simd_into(val.simd)
 }