prims.div inconsistencies between executors for integer tensors

[nikitaved]

Likely the cause of #808

In [1]: import torch, thunder

In [2]: def f(x, y):
   ...:     return thunder.prims.div(x, y)
   ...: 

In [3]: x = torch.randint(1, 10, (3, 3), device="cuda")

In [4]: y = torch.randint(1, 10, (3, 3), device="cuda")

In [5]: jf = thunder.jit(f)

In [6]: jf(x.cpu(), y.cpu())
Out[6]: 
tensor([[0, 1, 1],
        [0, 0, 2],
        [0, 0, 0]])

In [7]: jf(x, y)
Out[7]: 
tensor([[0.2500, 1.3333, 1.0000],
        [0.3333, 0.1250, 2.2500],
        [0.1429, 0.3750, 0.5000]], device='cuda:0')

In [8]: trace = thunder.last_traces(jf)[-1]

In [9]: print(trace)
# Constructed by Delete Last Used (took 0 milliseconds)
import torch
from thunder.executors.torchex import no_autocast

@torch.no_grad()
@no_autocast
def computation(x, y):
  # x: "cuda:0 i64[3, 3]"
  # y: "cuda:0 i64[3, 3]"
  [t0] = nvFusion0(x, y)
    # t0 = prims.div(x, y)  # t0: "cuda:0 i64[3, 3]"
  del x, y
  return t0

cc @apaz-cli