Document how to use generated Cython `.pxi` headers

Nobody has claimed this yet.

Sometimes when compiling pandas you end up having to compile twice to get the library to completely build. The files that get compiled the second time tend to be the outputs of a Tempita process.

Here are some small excerpts from pandas (some parts intentionally omitted to try and keep minimal):

```meson
_hashtable_class_helper = custom_target('hashtable_class_helper_pxi',
output: 'hashtable_class_helper.pxi',
input: 'hashtable_class_helper.pxi.in',
command: [
py, tempita, '@INPUT@', '-o', '@OUTDIR@'
]
)
_hashtable_func_helper = custom_target('hashtable_func_helper_pxi',
output: 'hashtable_func_helper.pxi',
input: 'hashtable_func_helper.pxi.in',
command: [
py, tempita, '@INPUT@', '-o', '@OUTDIR@'
]
)

...
cython_args = [
'--include-dir',
meson.current_build_dir(),
'-X always_allow_keywords=true'
]

py.extension_module(
hashtable,
sources: ['hashtable.pyx', _hashtable_class_helper, _hashtable_func_helper],
cython_args: cython_args,
...
)
```

When I look at the dotgraph that ninja generates, I noticed that the tempita outputs are declared as dependencies of both the `cython_COMPILER` and `c_COMPILER` steps, which is where I think a possible race condition leading up to the `c_COMPILER` step could be what requires a recompile.

Shouldn't the generated cython files only be a dependency for the `cython_COMPILER`? Is there a way to explicitly declare this?

Start with the Meson custom_target and py.extension_module excerpts in the issue, then inspect the generated Ninja dotgraph behavior described there. Determine whether the generated .pxi files should depend only on the Cython compiler and document how to declare that dependency explicitly, including how to verify the resulting build graph.

Written by the indexing model from the issue text.