support inferring dtype with torch.get_default_dtype for factory functions

[kshitij12345]

Fixes: #750

Changes -

1. Stash torch.get_default_dtype in cache_info. Also, this adds a check to the prologue trace to verify that jitted fn is called with same default dtype - see example prologue below.
2. Factory functions infer the dtype based on torch.get_default_dtype from cache_info (if dtype is not passed explicitly)
3. We don't support changing the default dtype in the jitted fn as reordering and fusion can lead to issue - it is a loud error for now (we can revisit in follow-up if required).

Repro:

import torch
import thunder

def foo(x: torch.Tensor) -> torch.Tensor:
    o = torch.zeros(x.shape, device=x.device)
    return o

jfoo = thunder.jit(foo)
o = jfoo(torch.randn(3, 3))
print(o.dtype)
print(thunder.last_prologue_traces(jfoo)[0])
print()
print(thunder.last_traces(jfoo)[0])

Prologue Trace

import thunder
import thunder.core.prims as prims
import torch
from thunder.executors.torchex import no_autocast

@torch.no_grad()
@no_autocast
def prologue(*args, **kwargs):
  # args: "Any"
  prims.check_len(args, 1)
  # kwargs: "Any"
  prims.check_len(kwargs, 0)
  t_0: "cpu f32[3, 3]" = args[0]
  prims.check_tensor_metadata(t_0, (3, 3), 'cpu', torch.float32, False)
  cache_info: "Any" = thunder._get_cache_info()
  cache_info_default_dtype: "<class 'torch.dtype'>" = cache_info['default_dtype']
  # NOTE - We bake the torch.dtype in trace (check_tensor_metadata also does the same).
  prims.check_literal_like(cache_info_default_dtype, torch.float32)
  cache_info_is_autocast_enabled: "bool False" = cache_info['is_autocast_enabled']
  prims.check_number_type_and_value(cache_info_is_autocast_enabled, False)
  cache_info_no_grad_sync: "bool False" = cache_info['no_grad_sync']
  prims.check_number_type_and_value(cache_info_no_grad_sync, False)
  return ((), ())

Computation Trace

import thunder
import thunder.core.devices as devices
import thunder.torch as ltorch
import torch
from thunder.executors.torchex import no_autocast

@torch.no_grad()
@no_autocast
def computation():
  # /home/kkalambarkar/lightning-thunder/scratchpad/test.py:64:             o = torch.zeros(x.shape, device=x.device)
  o = ltorch.zeros((3, 3), device=devices.Device("cpu"), dtype=None)  # o: "cpu f32[3, 3]"
    # o = ltorch.full((3, 3), 0, device=devices.Device("cpu"), dtype=None)  # o: "cpu f32[3, 3]"
      # o = prims.full((3, 3), 0, device=devices.Device("cpu"), dtype=dtypes.float32)  # o: "cpu f32[3, 3]"
  return o

[t-vi]

Supergood, thank you @kshitij12345