[Enhancements] (Much Better) Loop Fusion and Packed Funcall Mutation (for unrolling/vectorizing/tiling).

[hikettei]

TODO in this PR

• Remove duplicated index computation process by using expr-eq
  • If tmpvar is only used in the kernel; don't allocate, use scalar instead.
• Implement Loop Collapse in transform.lisp level. (judge the validity based on permutable and coincidence)
• Implement Loop Unrolling in transform.lisp (Packed Funcall)
• Implement Symbolic Loop Unrolling by introducing loop-reminder.
• Unroll the outermost iteration
• Test Unrolling
• Pass gemm-test (by fixing invalid shape inference)
• Fix: !rand compilation segv
• !softmax in a single kernel
• Fix for !sum !mean accuracy in CI. (Added loop fusion patterns)
• Fix: potentially there are still invalid RaW/WaW/WaR
• Fix: threefry2x32 in two kernels.
• Eliminate guard node (e.g.: c0 == 0)
  • as well as symbolic mode
  • Appeared in fused !tan, LayerNorm, and ConvND
• float4 accumlation
• optimize for :reduction; create float _acc_0
• Optimize for metal: use float4
• Compiling !rand in 10s.
• Simplifier: Fold Scalars
• Make Symbolic Compilation completely equivalent to static one.
  • Involve shape computation in a VM level op.
• BEAM Search for determining global_size/local_size
• Lisp JIT
• Fuse Conv < 2 Kernels
  • delete guard: c0==0
  • relocated_to_most_nearest
• Add NN Ops test with restricted accuracy
  • make all tests green
• BugFix:

(let ((a (make-tensor `(5 5))))
	(caten (!add (!view a `(0 2) `(0 2)) (!view a `(2 4) `(2 4)))))