Avoid PackTranspose calls Fp32 tiny blas kernels

[shalinib-ibm]

This patch gets rid of the redundant vec_perm insns by re-ordering the matrix multiplication in kernel

Make sure to read the contributing guidelines before submitting a PR

[shalinib-ibm]

llama_file sgemm computes C = A^T * B in column major
For MMA , this equates to C^T = A * B^T in row major
Now A * B^T ( in row major) = (B * A^T)^T where B and A^T should be row major, but we have B and A^T in column major order only
So, this approach is not working

[shalinib-ibm]

MMA example with 4x4 matrix multiplication : 

A (row-major)     => 4×4:

[ [1, 2, 3, 4],

  [5, 6, 7, 8],

  [9,10,11,12],

  [13,14,15,16] ]

B (row-major)     => 4×4:

[ [17,18,19,20],

  [21,22,23,24],

  [25,26,27,28],

  [29,30,31,32] ]

Optimal MMA setup :
Transpose matrix A to get columns of A as rows : 

Compute C = A^T times B

xvf32gerpp(acc, col0_of_A, row0_of_B);

xvf32gerpp(acc, col1_of_A, row1_of_B);

xvf32gerpp(acc, col2_of_A, row2_of_B);

xvf32gerpp(acc, col3_of_A, row3_of_B);

Problem with llamafile_sgemm

A^T (column-major layout):

Memory: [1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16]

B (column-major layout):

Memory: [17 21 25 29 18 22 26 30 19 23 27 31 20 24 28 32]

Indexes:

I	vec_A row	vec_B row	Indices
0	[1, 5, 9, 13]	[17, 18, 19, 20]	0,4,8,12
1	[2,6,10,14]	[21, 22, 23, 24]
2	[3,7,11,15]	[25,26,27,28]
3	[4,8,12,16]	[29,30,31,32]

Row-major Aᵀ (i.e., A stored row-major, Aᵀ read as rows):
You can use vec_xl(0, A + l + x * 4) — clean 16-byte contiguous loads.

Column-major Aᵀ (A stored column-major, same memory order but semantics changed):
Now you need non-contiguous access (strided by 4 floats) to gather one row of Aᵀ (which is one column of A).
That requires 4 separate loads or a vector gather