Fix syntax error

Ensure that the caller and target processes have matching architectures
before proceeding. Attaching to a 64-bit process from a 32-bit process
will fail when using CreateToolhelp32Snapshot, and attempting to attach
to a 32-bit process from a 64-bit process may cause issues during PE
file parsing. To avoid potential errors abort the operation if the
architectures differ.

This reverts commit 746ecfc.

That commit isn't correct. It conflates the return from `IsWow64Process`
(whether the process is running under 32-bit emulation on a 64-bit
process) with whether the process is 64-bit. A 32-bit process on 32-bit
Windows would have `IsWow64Process` set our `isWow64` flag to `FALSE`,
and we'd then incorrectly return `TRUE` from `is_process64Bit`.

…_support

External debugger Windows support

This is useful because debuggers will write to this variable remotely,
and it's helpful for it to have a well known size rather than one that
could vary per platform.

Require the flags for turning this on to be set to exactly 1 to avoid
accidentally triggering remote debugging in the case of heap corruption.
Make a heap copy of the script path before using it to avoid the buffer
being overwritten while we're still using it by another debugger.

This avoids the need to reopen it by (wide character) path.

Previously it was leaked if `PyObject_AsFileDescriptor` failed.

I think this was a copy paste error.

The remote process must be a compatible version.

Explain our compatibility rules.

We can clean this up by freeing strings as soon as we're done with them,
and remove some duplicate code.

This was documented as working, but in fact we assumed that the path
would always be a unicode string. Factor the version that only accepts a
unicode string into a helper function, and have the caller call
`os.fsdecode` on the input path, which accepts either a byte string or a
unicode string and always returns a unicode string.

Previously we were returning the number of bytes, which we didn't need
and would never be different than the argument that was passed into the
function.

This field in the structure is defined as an `unsigned long`, not as
a `pid_t`.

We already did a `strlen` on the `debugger_script_path`, so it's too late
to check if it's NULL here!

This helps to harden a bit against heap corruption.