Support slice for device code

Revival of https://github.com/rust-accel/accel/pull/16

Objective

Slice patterns in the kernel arguments become to work, and make it safe function

#[kernel]
pub fn add(a: &[f64], b: &[f64], c: &mut [f64]) {
    let i = accel_core::index() as usize;
    if i < c.len() {
        c[i] = a[i] + b[i];
    }
}

Semantics

a, b, c and other slices are passed as a element length (in size, not byte) and a pointer of actual data. Pointer must be one of following:

• a device pointer (i.e. casted from CUdeviceptr)
• a host pointer which can be accessed from device directly.
  • e.g. a host memory allocated by cuMemAllocHost

TODO

• Add custom trait for device sendable to change its behavior for primitive types / slice
  • DeviceSend trait in !41 (merged)
• Implement DeviceSend to &[T] and &mut [T]
• Test slice can be access from device code