Changes

New backend: aten, aten[clang]
- It entirely relies on the tiny JIT Compiler (called AbstractTensor.lisp)
  - It is tiny and easy to write a new backend.
  - It optimises the kernel, especially where memory bandwidth is an overhead. (Conv2d gets approximately 10 times faster on CPU.)
  - This method (could be) extended to GPU architectures.
  - [ ] (WIP) It unrolls/vectorizes the kernel, enabling SLEEF/Neon/AVX intrinsics.
  - [ ] (TBD) data-parallel computation using OpenMP.
- Now it supports basic operations including arithmetic ops and mathematical ops (not tested).
- (TODO) Implement Gemm/Im2Col/Winograd etc.
Removed a JITCPUTensor backend.
Fixed a bug related to the VM execution model. %vm-move ignores the broadcast. (resulting the shape error.)

Usage

e.g.:

(with-devices ((Aten[Clang] :debug 1 :opt 3) CPUTensor)
    ....)

$ ./roswell/waffe2.ros test -b "Aten[Clang]" -b LispTensor -b CPUTensor