BUG: Segfault constructing pd.Timedelta with numpy 2.3.x/2.4.x on Python 3.14 (pandas 3.0.4)

### Pandas version checks

- [x] I have checked that this issue has not already been reported.

- [x] I have confirmed this bug exists on the latest version of pandas (3.0.4).

### Reproducible Example

```python
import pandas as pd
pd.Timedelta(1, unit="D")   # interpreter exits with SIGSEGV (signal 11)
```

`import pandas` succeeds; `numpy.timedelta64(1, "D")` succeeds; the **first**
`pd.Timedelta(...)` construction crashes the interpreter. No Python traceback is
produced.

### Issue Description

On Python 3.14, `pandas==3.0.4` segfaults when constructing a `Timedelta` if the
installed numpy is in the 2.3.x–2.4.x range. The crash is a hard SIGSEGV
(exit 139), single-process and fully deterministic.

The numpy `_multiarray_umath` shared object is not stripped, so the faulting
frame symbolizes exactly:

```
#0  PyUFunc_f_f (numpy/_core/_multiarray_umath.cpython-314-x86_64-linux-gnu.so + 0xd5f58)
        => mov (%rdx),%r15            # registers at fault: rdx = 0x6, rcx = 0x0
#1  pandas/_libs/tslibs/timedeltas.cpython-314-x86_64-linux-gnu.so
#2  pandas/_libs/tslibs/timedeltas.cpython-314-x86_64-linux-gnu.so
#3  slot_tp_new -> type_call -> _PyObject_MakeTpCall    # Timedelta(1, unit='D')
```

`PyUFunc_f_f` is numpy's generic float32 unary ufunc inner loop, whose signature
is `(char **args, npy_intp *dimensions, npy_intp *steps, void *func)`. At the
fault, `rdx` (the `steps` pointer) holds `0x6` and `rcx` (`func`) is `NULL`; the
loop dereferences `steps` and segfaults.

I bisected across numpy/pandas versions, each in a clean venv (Python 3.14.5),
running only the one-liner above:

| numpy  | pandas | result  |
|--------|--------|---------|
| 2.3.5  | 3.0.3  | OK      |
| 2.3.5  | 3.0.4  | SIGSEGV |
| 2.4.6  | 3.0.4  | SIGSEGV |
| 2.5.0  | 3.0.4  | OK      |

Observations (reporting only, not asserting a cause):

- pandas 3.0.3 and 3.0.4 both declare `numpy>=2.3.3; python_version >= "3.14"`.
  3.0.3 runs fine on numpy 2.3.5; 3.0.4 segfaults on the same numpy.
- 3.0.4 runs fine on numpy 2.5.0 but segfaults on numpy 2.3.5 and 2.4.6.
- Reproduced on three different x86-64 CPUs/hosts (one with AVX2, one without)
  and inside a `python3.14` Debian-bookworm container — so it does not appear to
  be CPU- or host-specific.

### Expected Behavior

`pd.Timedelta(1, unit="D")` returns a `Timedelta` (or raises a Python exception)
rather than segfaulting, for any numpy version permitted by pandas's declared
dependency metadata.

### Installed Versions

<details>

```
python : 3.14.5 (x86_64, glibc 2.42)
pandas : 3.0.4   (manylinux x86_64 wheel, PyPI)
numpy  : 2.3.5   (manylinux x86_64 wheel, PyPI)
OS     : Linux 7.0.12 (Fedora 43); also reproduced on Debian bookworm container
install: clean venv -> `pip install numpy==2.3.5 pandas==3.0.4`
```

</details>

[]