simplify matrix vector multiplication

[Traxar]

changes

• reduce the fields into an array of columns
• use simd to calculate all 4 entries at once
  • use a inline loop to further reduce the sum
• remove the now unnecessary row function

[Traxar]

packedness is lost due to using an array, which does not guarantee memory layout. The sample still works (atleast for me). If packedness is of importance i will rewrite the changes both cpu and gpu side with the original struct fields and the simplified maths

[Gota7]

Hm now this is a good question. If we mark the struct as extern that conforms to the C ABI packing right? But is that consistent across platforms?

[Traxar]

i think the extern mark should work 👍