Polars assumes microseconds instead of reading numpy timedelta units

[hixan]

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 Polars.

Reproducible example

import numpy as np, numpy as np
print(pl.DataFrame([[np.timedelta64(1, 'ns')]]))
>>> shape: (1, 1)
┌──────────────┐
│ column_0     │
│ ---          │
│ duration[μs] │
╞══════════════╡
│ 1µs          │
└──────────────┘

Log output

No response

Issue description

Polars does not seem to read the time unit associated with the numpy timedelta64, and instead assumes microsecond.

Expected behavior

The returned column should have the same time unit as the numpy object (in this case ns)

Installed versions

``` --------Version info--------- Polars: 0.20.23 Index type: UInt32 Platform: Linux-5.14.0-284.18.1.el9_2.x86_64-x86_64-with-glibc2.28 Python: 3.11.8 (main, Feb 16 2024, 19:42:16) [GCC 8.5.0 20210514 (Red Hat 8.5.0-20)] ----Optional dependencies---- adbc_driver_manager: cloudpickle: 3.0.0 connectorx: deltalake: fastexcel: fsspec: 2024.3.1 gevent: hvplot: matplotlib: 3.8.4 nest_asyncio: 1.6.0 numpy: 1.26.4 openpyxl: 3.1.1 pandas: 1.5.3 pyarrow: 11.0.0 pydantic: 1.10.15 pyiceberg: pyxlsb: sqlalchemy: 1.4.52 xlsx2csv: xlsxwriter: ```

[datenzauberai]

Handling of time units and proper casting is missing when the numpy values are inside a python list instead of a numpy arrray:

pl.DataFrame({
    # ok
    "np_array_ns": np.array([np.timedelta64(1000000000, "ns"), np.timedelta64(1000000000, "ns")]),
    # ok
    "np_array_us": np.array([np.timedelta64(1000000, "us"),    np.timedelta64(1000000, "us")]),
    # ok
    "np_array_ms": np.array([np.timedelta64(1000, "ms"),    np.timedelta64(1000, "ms")]),
    # ignores time unit
    "py_list_ns": [np.timedelta64(1000000000, "ns"), np.timedelta64(1000000000, "ns")],
    # TypeError
    "py_list_us": [np.timedelta64(1000000, "us"),    np.timedelta64(1000000, "us")],
    # TypeError
    "py_list_ms": [np.timedelta64(1000, "ms"),    np.timedelta64(1000, "ms")],
})