Packed types support {u,i}16, {u,i}8

[andreytkachenko]

Dealing with packed types may be much better, ideally like Tensor<u16> support, with implicit packing and unpacking, alternative variant - is a packed types with pack/unpack methods, like:

• u16x2 - packed: [u16, u16], unpacked to (u32, u32)
• i16x2 - packed: [i16, i16], unpacked to (i32, i32)
• u16x2f - packed: [u16, u16], unpacked to (f32, f32) with range 0..1
• i16x2f - packed: [i16, i16], unpacked to (f32, f32) with range -1..1
• u8x4 - packed: [u8, u8, u8, u8], unpacked to (u32, u32, u32, u32)
• i8x4 - packed: [i8, i8, i8, i8], unpacked to (i32, i32, i32, i32)
• u8x4f - packed: [u8, u8, u8, u8], unpacked to (f32, f32, f32, f32) with range 0..1
• i8x4f - packed: [i8, i8, i8, i8], unpacked to (f32, f32, f32, f32) with range -1..1

[nathanielsimard]

Not a bad idea. For now, we would have Tensor<Line<u16>>, but it's a long type. Maybe we should hardcode the fact that tensors always return Line<I>. You could access the line length with tensor.line_length(). So the minimum line length for Tensor<u16> would be 2, but it could be more depending on the launch options.