internal ref for why its needed https://fb.workplace.com/groups/1075192433118967/permalink/2031263990845135/
Adds `min` / `max` kwargs to `torch._dynamo.maybe_mark_dynamic`, so a dim can be marked
dynamic *and* given an initial range without enforcing that the dim stays dynamic:

```python
torch._dynamo.maybe_mark_dynamic(x, 0, min=2, max=5)
```
The range is a starting point: guards may narrow it ([2, 5] + a x.shape[0] < 5 guard compiles with (2, 4)), and the dim may still specialize instead of raising, unlike mark_dynamic.

A hint outside the declared range is still an error (ConstraintViolationError: 6 not in range [2, 5]) — the range is never widened to fit the input.

Bugs fixed along the way
- min or max alone crashed with AssertionError: not simple sympy type NoneType (ValueRanges rejects None). The missing bound now defaults to 2 / int_oo, the bounds a backed size already has. This also fixes mark_dynamic(x, 0, min=2).
- A dim made dynamic by AOTAutograd's _dynamo_propagated_dynamic_indices, on a tensor carrying a range for another dim, raised IndexError: pop from empty list.

- changing min/max between calls now recompiles rather than reusing the cached graph, and partial ranges no longer crash.
- Marking one dim with both mark_dynamic and maybe_mark_dynamic now raises; different dims of the same tensor may still use different APIs.

Pull Request resolved: https://github.com/pytorch/pytorch/pull/194091
Approved by: https://github.com/bobrenjc93