2x Times Faster Training by various optimizations

[hikettei]

Changes

!move/!copy ... :maybe-in-place

• With set to T, MoveTensorNode is replaced with ALLOC{INTERNAL}.

CL-WAFFE2-REPL> (disassemble-waffe2-ir (!sin (!sin (randn `(3 3)))))

disassemble-waffe2-ir:
 [Forward]: 
<WfInst[op=ALLOC{INTERNAL}]   : TID40883 <= op(TID40883{float, (3 3)} <Input>TID40880{float, (3 3)})>
<WfInst[op=SINNODE-CPUTENSOR] : TID40883 <= op(<Input>TID40880{float, (3 3)} TID40883{float, (3 3)})>
<WfInst[op=SINNODE-CPUTENSOR] : TID40883 <= op(TID40883{float, (3 3)} TID40883{float, (3 3)})>

3 Instructions | 2 Tensors | 0 Scalars

 [Pullback]: 

0 Instructions | 0 Tensors | 0 Scalars

BugFix: Shape-Error

• Tensors in error-free positions are now also displayed.

call-with-view

• [Fix] With JIT, SIMD registers are not fully used because Loop Collapse is disabled. Now it works
• [Fix] Loop Collapse/and so on is enabled whenever compiled with defmodel-as