Post-MultiExpr Fusion in the equivalent domain

[hikettei]

• Post-MultiExpr (Completely Equivalent Domain)
• Post-MultiExpr (where subdomain is the equivalent)
• Index-Component-Fusion
• Add the following transform pattern to EXPR:
• More Restrict Schedule test in nn or ajit (e.g.: Embedding in one kernel)
• matmul transpose fusion by multiexpr? (subdomain)
• fix for double free (?) (rng-counter might be related to this)

val_445[(_gid0+0)] = (val_456_0^((val_445[(_gid0+0)]*32768)+(uint32_t)((float)val_445[(_gid0+0)]*7.6293945e-6)));
->
float _tmp_0 = val_445[_gid0+0];
val_445[_gid0+0] = (val_456_0^((_tmp_0*32768)+(uint32_t)((float)_tmp_0*7.6293945e-6)));

• Fix: 6d transposed matmul scheduler

• randn < 2 kernels including scalar parts

  • Tests for In-place Embedding, 2 Kernel Randn
  • Accuracy tests for RMSNorm, Embedding, etc ...

• Deleting the dead code:

void main500811_e229_k0(uint32_t* val_400, uint32_t* val_123, uint32_t val_397) {
  (*val_123) = 4;
  (*val_400) = (val_397+(*val_123));
  uint32_t val_124 = (*val_400);
}