Precalculate float consts for RecursiveGaussian at build time

[FreezyLemon]

RecursiveGaussian::new does calculations that are always based on the same constant values defined at compile-time. My first thought was to just inline these, but that would make maintaining the algorithm a lot more difficult.

This PR adds a simple build script that does the same as RecursiveGaussian::new and writes the resulting values to const variables:

mod consts {
  pub const RADIUS: usize = 5_usize;
  pub const VERT_MUL_IN_1: f32 = 0.055295236_f32;
  pub const VERT_MUL_IN_3: f32 = -0.058836687_f32;
  pub const VERT_MUL_IN_5: f32 = 0.012955819_f32;
  pub const VERT_MUL_PREV_1: f32 = -1.9021131_f32;
  pub const VERT_MUL_PREV_3: f32 = -1.1755705_f32;
  pub const VERT_MUL_PREV_5: f32 = -0.00000000000000012246469_f32;
  pub const MUL_IN_1: f32 = 0.055295236_f32;
  pub const MUL_IN_3: f32 = -0.058836687_f32;
  pub const MUL_IN_5: f32 = 0.012955819_f32;
  pub const MUL_PREV_1: f32 = 1.9021131_f32;
  pub const MUL_PREV_3: f32 = 1.1755705_f32;
  pub const MUL_PREV_5: f32 = 0.00000000000000012246469_f32;
  pub const MUL_PREV2_1: f32 = -1_f32;
  pub const MUL_PREV2_3: f32 = -1_f32;
  pub const MUL_PREV2_5: f32 = -1_f32;
}

The RecursiveGaussian can then use these consts instead of self.asdf in the blur algorithm.

I understand that there is a reason that float calculations are not const in Rust (even though integer calculations are), but by my testing, even platforms without FMA produce the exact same variable values. Maybe more testing is needed on ARM and maybe Windows platforms (due to their x87 FPU shenanigans that might apply here). But I think that even if there are inconsistencies, they're probably smaller than the inaccuracies the algorithm itself would produce at runtime anyways.

I have no clue why I never tested this before. The results are staggering for the amount of effort needed:

blur                    time:   [9.0900 ms 9.1005 ms 9.1112 ms]
                        change: [-36.050% -35.935% -35.820%] (p = 0.00 < 0.05)
                        Performance has improved.

This also reduces the resulting library size (very roughly 20%), mostly because of nalgebra being removed from runtime.

[FreezyLemon]

Oops, still some clippy lints to sort out