With MSVC the Ninja generator extracts the `cl -showIncludes` prefix.
When MSVC is configured to have non-English output, e.g. via
`VSLANG=2052` in the environment, then `cl` prints the prefix encoded
for the current code page, which is not necessarily UTF-8 encoding.
Currently we fail to convert the prefix to our internal UTF-8 encoding,
but assume it is UTF-8 later.

While writing `rules.ninja`, the Ninja generator converts our internal
UTF-8 encoding to the current code page.  The `msvc_deps_prefix =` line
needs to be encoded as the current code page so that `ninja` can match
in the output from `cl -showIncludes` during the build.

Prior to commit v3.9.0-rc1~47^2 (codecvt: Re-implement do_out and
do_unshift, 2017-05-25), the non-UTF-8 prefix extracted above was
written without noticing its incorrect internal encoding.  The
`rules.ninja` file was successfully written, but possibly with a mangled
`msvc_deps_prefix`.  Since that commit the output stream correctly
rejects the non-UTF-8 byte sequence and writing `rules.ninja` fails.

Fix this by correctly converting the `cl -showIncludes` output from the
current code page to our internal UTF-8 encoding.

Fixes: #17191