Compile entire crate by nvptx64-nvidia-cuda target

[termoshtt]

A proposition to resolve #61

Motivation

#[kernel] function cannot use any variable, function, and so on because it will be compiled as a stand alone device code.

fn add_2(a: &mut f32) {
    *a = *a + 2.0;
}

#[kernel] 
pub fn add_2_all(a: *mut f32, n: usize) {
    let i = ::accel_core::index();
    unsafe { add_2(&mut *a.offset(i)) };  // add_2 cannot find
}

Resolution

Compile entire crate both as x86_64 and nvptx64 targets.

• rust-ptx-linker will eliminate non-PTX kernel code which does not called from PTX kernel

Problems

• [ ] std must be compiled with nvptx
• [ ] Compile flow (How to trigger nvptx build instead of proc-macro?)