Metal: Metal ignores default Matrix layout options of Slang

[ArielG-NV]

Related to: #4291 tests/compute/column-major.slang test produces incorrect results. The results seem to hint that the matrix used is not swapping its major'ness.

The emitting-code seems to reflect this, getTargetProgram()->getOptionSet().getMatrixLayoutMode() is not taken-into-account when emitting metal code.

[ArielG-NV]

Requirements:

• new pass 'lowerColumnMajor' - done
• add lowerColumnMajor pass which metal runs (due to lacking column/row major attributes) - done
• flip all matrix<n,m> into matrix<m,n> - done
• reorder all vector-matrix multiplies - done
  • note: [1]
• matrix multiplies need to be reordered - done
  • note: [1]
• legalize matrix accessors (getElement / getElementPtr) into valid operations (construct a vector when fetching a column and indexing can work off of that, else index directly into matrix) - done
• reorder swizzles into valid accessors --> _mXY and _XY swizzles - done
• handle assignment of matrixes with different layouts. This should be handled through wrapping these types with a MatrixTypeCast op which is implemented accordingly - not done
• Make_Matrix
  • scalars not done
  • vectors not done
  • matrix of larger size not done
  • vector + matrix of larger size not done

[1] - we hack around mul currently (not all multiplies are intrinsic_op's, need to make them intrinsic_op's). Require to add the following:

• Add functionality/translation to emit'ing code - done
• Add auto-diff parts - not done

[ArielG-NV]

note: old method to solve the problem is being phased out for a new method which fixes and unifies the solution with GLSL/HLSL/SPIRV

1. fix packing not being respected by cpp/metal/cuda
2. lowerBufferElementTypeToStorageType() is where we should legalize matrix-load and matrix-store
3. remove all unneeded row_major/column_major modifier emits. Account the translation for the current set default layout (HLSL/GLSL/SPIRV). CPP/Metal/CUDA should always convert matrix layout from global to row_major since there is no default layout allowed for these targets.

// optional...

4. kIROp for transpose to legalize load/store?
5. autodiff transpose?