Function build_program_sharded 

pub fn build_program_sharded(
    workgroup_size_x: u32,
    opcodes: &[OpcodeHandler],
) -> Program

Build the megakernel IR with a custom workgroup size and optional custom opcodes.

Buffers are declared with concrete with_count(...) sizes so the backend readback layer allocates the right static staging size - a count=0 default reads back 4 bytes regardless of how much the kernel wrote.