Add full support for set_grad_enabled

[shino16]

Fixes #2398. Thunder has support for torch.set_grad_enabled calls that come from torch.no_grad() context by managing cd.is_grad_enabled without actually calling torch.set_grad_enabled. As in #2398, when Thunder's graph splitter splits the region, torch.set_grad_enabled(True/False) appears in each segment independently without its counterpart. We should actually call it, and it should be done before connecting the computation to Autograd so that it will populate the grad_fn attribute appropriately.

[shino16]

Currently, the following does not work:

import torch, thunder

torch.set_grad_enabled(False)
def fn(x):
    with torch.enable_grad():
        return x.sin()

x = torch.randn((), requires_grad=True, device="cuda")
assert fn(x).grad_fn is not None
assert thunder.jit(fn)(x).grad_fn is not None  # AssertionError
assert thunder.dynamo.thunderfx(fn)(x).grad_fn is not None  # AssertionError

This is because:

1. the epilogue calls unpack_trivial(t0), where t0 represents x.sin()
2. the output of unpack_trivial is also t0
3. autograd is disabled when tracing the epilogue
4. so the output of unpack_trivial is tagged DETACHED_AUTOGRAD_GRAPH meaning it will be treated as constant

This maybe a similar issue to #2396 (comment).

[shino16]

Let me close this PR as the purpose has somewhat changed. I'll reopen as a new PR.