Support for A5

[zouzias]

Keep track of A5 support on kernels.

• make compile_a5_abs
• make compile_a5_csr_gather
• make compile_a5_batch_matrix_square
• make compile_a5_chunk_cumsum
• make compile_a5_gdn_chunk_h
• make compile_a5_gdn_chunk_o
• make compile_a5_gdn_scaled_dot_kkt
• make compile_a5_gdn_wy_fast
• make compile_a5_scan_ul1
• make compile_a5_simple_matmul
• make compile_a5_swiglu
• make compile_a5_tri_inv_col_sweep
• make compile_a5_tri_inv_ns
• make compile_a5_tri_inv_rec_unroll
• make compile_a5_tri_inv_trick

Work in progress

Most AIV kernels fail due to error: the range of 1st parameter must be [4, 6] pipe_barrier(PIPE_V); .

[zouzias]

This looks a bit challenging to fix:

csrc/kernel/kernel_gdn_wy_fast.cpp:415:30: error: function type 'void (long) noexcept' of 'wait_flag_dev' does not support the given target feature
            if (!first_iter) wait_flag_dev(3);
                             ^
csrc/kernel/kernel_gdn_wy_fast.cpp:453:26: error: the range of 1st parameter must be [4, 6]
            pipe_barrier(PIPE_V);
                         ^
csrc/kernel/kernel_gdn_wy_fast.cpp:455:26: error: the range of 1st parameter must be [4, 6]
            pipe_barrier(PIPE_V);
                         ^
csrc/kernel/kernel_gdn_wy_fast.cpp:457:26: error: the range of 1st parameter must be [4, 6]
            pipe_barrier(PIPE_V);
                         ^
csrc/kernel/kernel_gdn_wy_fast.cpp:462:30: error: function type 'void (long) noexcept' of 'wait_flag_dev' does not support the given target feature
            if (!first_iter) wait_flag_dev(4);