[GENX]: Enhance operator `genx.matrix.dpas` to allow more input shapes

[etiotto]

The genx.matrix.dpas operator is lowered to a SIMT DPAS instruction. The operands of the DPAS instructions (A,B) have several constraints, depending on the data type of the vector element. These are the expected input operands shapes:

1. If the size of the element is 16 bits (f16, bf16): the expected input A, B operands' types are vector<8 x f16> or vector<8 x bf16>, the GENX translation generates a bitcast to vector<8 x i16> to satisfy IGC's requirements

2. If the size of the element is 32 bits (tf32): the expected input A, B operands types are vector<4 x tf32>, i.e., half of the row or column, the GENX translation generates a bitcast to vector<4 x i32>

3. If the size of element is 8 bits (i8, u8): the expected input A, B operands types are vector<8 x i16/u16>, i.e. 2 adjacent elements of i8/u8 are packed into a 16 bit element type

For case (1) there are no issues while for case (2) and (3) the operator requirements cause the producer to pack data before creating the operation, specifically:

• case 2: if the producer input types (for A, B) are vector<8 x tf32> it needs to split it to two vector<4 x tf32> between two SIMD lanes/Warp elements before calling the GENX operation.

• case 3: if the produces input types (for A, B) are vector<16 x i8/u8> it needs to pack the two adjacent 8 bit elements into a vector<8 x i16/u16> before calling the GENX operation

The genx.matrix.dpas operator should be improved to allow inputs of type vector<8 x tf32> (case 2) and vector<16 x i8/u8> (case 3). The operation should take care of packing or splitting the input vectors appropriately.