perf(vello_cpu): statically dispatch f32 image sampling quality

[tomcur]

By using a const generic to dispatch statically between bilinear and bicubic sampling, on x86 we get a 7% timing decrease for bilinear sampling (medium quality) in the f32 pipeline. It appears not to impact timings for bicubic sampling (high quality).

The benchmark was performed by temporarily adding medium and high quality variants of transform::rotate in
sparse_strips/vello_bench/src/fine/image.rs.

Benchmarking fine/image/transform/rotate_medium_f32_scalar: Collecting 100 samples in estimated 5.0177 s (495k itfine/image/transform/rotate_medium_f32_scalar
                        time:   [10.137 µs 10.144 µs 10.151 µs]
                        change: [-7.5548% -7.4067% -7.2867%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 13 outliers among 100 measurements (13.00%)
  2 (2.00%) low mild
  5 (5.00%) high mild
  6 (6.00%) high severe

Benchmarking fine/image/transform/rotate_high_f32_scalar: Collecting 100 samples in estimated 5.0365 s (162k iterfine/image/transform/rotate_high_f32_scalar
                        time:   [31.163 µs 31.293 µs 31.556 µs]
                        change: [-0.1547% +0.4795% +1.2307%] (p = 0.19 > 0.05)
                        No change in performance detected.
Found 13 outliers among 100 measurements (13.00%)
  3 (3.00%) high mild
  10 (10.00%) high severe