[WIP] polyhedral compiler passing all tests

[hikettei]

• Implement Backward
• %where compilation fails
• Fix tan (needs to implement in-place mutation)
• Integrate all Expr renderers into %render-expr
• Affine Scheduling

ばぐ〜〜

• tmpvar 宣言して使わない
• Recursively Groupingの範囲 vs VMで計算する範囲？
• reductionのraw dependencies?

[hikettei]

Softmax fused in a single kernel:
(schedule_count needs to be included in the test)

TEST> (caten (!softmax (make-tensor `(3 3) :initial-element 1.0)))
Compiled:

#include <math.h>
#include <stdint.h>
#define boolean _Bool
#define min(a, b) ((a) < (b) ? (a) : (b))
#define max(a, b) ((a) > (b) ? (a) : (b))
/*
Arrays:
  - val_6[float32]: (3 1)
  - val_0[float32]: (3 3)
*/
void main1599824(float* val_6, float* val_0);
void main1599824(float* val_6, float* val_0) {
  float val_2;
  for(int c0=0;(c0<=2);c0+=1) {
    val_6[(c0)+(0)] = 0.0;
    for(int c1=0;(c1<=2);c1+=1) {
      val_2 = 1.442695;
      val_0[3*(c0)+(c1)] = 1.0;
      val_0[3*(c0)+(c1)] = val_0[3*(c0)+(c1)]*val_2;
      val_0[3*(c0)+(c1)] = exp2(val_0[3*(c0)+(c1)]);
      val_6[(c0)+0] += val_0[3*(c0)+(c1)];
    }
    val_6[(c0)+0] = 1.0 / (val_6[(c0)+0]);
    for(int c1=2;(c1<=4);c1+=1) {
      val_0[3*(c0)+((c1-2))] = val_0[3*(c0)+((c1-2))]*val_6[(c0)+0];
    }
  }
}

[hikettei]

TODO整理してMultiExpr直したらMergeする