pygmt.which: Errors if downloading multiple tiled grids

The issue was originally reported in https://github.com/GenericMappingTools/pygmt/pull/3148#issuecomment-2054814923.

Here is a minimal example to reproduce the issue.
```
!rm -r ~/.gmt/server/earth/earth_relief/earth_relief_15s_p # Make sure files are not downloaded
from pygmt import which
which(fname=["@N30W120.earth_relief_15s_p.nc", "@N00E000.earth_relief_15s_p.nc"], download="a")
```

The errors are:
```
File ~/OSS/gmt/pygmt/pygmt/src/which.py:67, in which(fname, **kwargs)
62 with lib.virtualfile_out(kind="dataset") as vouttbl:
63 lib.call_module(
64 module="which",
65 args=build_arg_string(kwargs, infile=fname, outfile=vouttbl),
66 )
---> 67 paths = lib.virtualfile_to_dataset(vfname=vouttbl, output_type="strings")
69 match paths.size:
70 case 0:

File ~/OSS/gmt/pygmt/pygmt/clib/session.py:1940, in Session.virtualfile_to_dataset(self, vfname, output_type, column_names, dtype, index_col)
1937 result = self.read_virtualfile(vfname, kind="dataset").contents
1939 if output_type == "strings": # strings output
-> 1940 return result.to_strings()
1942 result = result.to_dataframe(
1943 column_names=column_names, dtype=dtype, index_col=index_col
1944 )
1945 if output_type == "numpy": # numpy.ndarray output

File ~/OSS/gmt/pygmt/pygmt/datatypes/dataset.py:156, in _GMT_DATASET.to_strings(self)
154 if segment.contents.text:
155 textvector.extend(segment.contents.text[: segment.contents.n_rows])
--> 156 return np.char.decode(textvector) if textvector else np.array([], dtype=str)

File ~/opt/miniconda/envs/pygmt/lib/python3.12/site-packages/numpy/core/defchararray.py:615, in decode(a, encoding, errors)
572 @array_function_dispatch(_code_dispatcher)
573 def decode(a, encoding=None, errors=None):
574 r"""
575 Calls ``bytes.decode`` element-wise.
576
(...)
612
613 """
614 return _to_bytes_or_str_array(
--> 615 _vec_string(a, object_, 'decode', _clean_args(encoding, errors)))

TypeError: string operation on non-string array
```
The bug is most likely caused by an upstream API bug (for example, the memory that holds the text is mistakenly freed before writing), so it's not trivial to fix it.