Various mathematical transformations

[therontarigo]

Generally an improvement:

• cross(-a,b) or cross(a,-b) or -cross(a,b) -> cross(b,a) (related issue: sign cancellation)
• Left vs right multiplication:
  transpose(mat)*vec -> vec*mat
vec*transpose(mat) -> mat*vec
• transpose(a)*transpose(b) -> transpose(b*a)
• distance(a,b) -> length(a-b) (completed in 40dae3c)

Only if there's a flag to turn it off (otherwise could hurt compression):

• sqrt(dot(a,a)) -> length(a)
• min/max simplification (related issue: sign cancellation)
  min(-a,-b) -> -max(a,b)
-max(a,-b) -> min(-a,b)
  and so-forth.
• Common cross-product expansions:
  cross(vec3(1,0,0),a) -> vec3(0,-a.z,a.y)
cross(vec3(0,1,0),a) -> vec3(a.z,0,-a.x)
cross(vec3(0,0,1),a) -> vec3(-a.y,a.x,0)
  ( perhaps generalized to cross(vec3(d,0,0),a) -> d*vec3(0,-a.z,a.y) etc. (when d is known to be scalar type), though this doesn't seem common )

Expected impact on precision: none, driver shader compilers typically perform similar transformations anyway.