Bazel CI / TensorFlow: fix more errors (#9762)

* Bazel CI / TensorFlow: fix more errors

In this commit:

1. Keep just the first character of the python
version checker script's output. This is a simple
way to strip \n and \r and we only need the major
version anyway. The string is later used in
writing the .bazelrc file where these characters
would cause problems.

2. Export environment variables in
run_configure_for_cpu_build, use their default
values

3. Allow a user-defined value for BAZEL_SH

4. Fix python_configure.bzl:

4.a. Fix the symlink creation on Windows: create a
single junction pointing to the src_dir. This is a
trick to get around the problem that on Windows we
can't create symlinks. By creating this symlink
and declaring the same outputs as on Linux/Darwin,
we make it look like that the genrule successfully
produced these files. This only works in a
single-machine execution environment.

4.b. Get rid of the "select" in
@local_config_python//:numpy_include and
python_include. The results of `find` and `dir`
only make sense on Linux/Darwin and Windows
respectively, so there's no point in generating
the non-Windows genrule on Windows and the
Windows-specific one on non-Windows, because all
the paths would be wrong. So instead of that,
generate just the genrule corresponding to the
platform.

4.c. Write the genrule.cmd's commands to scripts
and execute those instead. This avoids the
difficulties with escaping backslashes and quotes.
Backslashes are necessary on Windows because paths
with forward slashes aren't generally accepted,
e.g. "cmd.exe /c foo/bar" doesn't work because
Windows believes foo is an executable and /bar is
its switch:

    c:\tempdir>type bar.cmd
    @echo this is bar.cmd, %%1=(%1)

    c:\tempdir>cmd.exe /c bar/baz
    this is bar.cmd, %1=(/baz)

5. simplify the "is_windows" function in
./configure

Fixes https://github.com/bazelbuild/bazel/issues/2892

* Fix bad cc_library.includes from prev. commit

* bazel_test_lib.sh: add envvars to CPU build config

* Fix junction creation in python_configure.bzl

Bazel deletes the output files from a previous
build before running an action, and if the
junction pointing to the source directory is
present then Bazel would delete the source files,
believing they are in the output directory.

Thanks to @meteorcloudy on
https://github.com/tensorflow/tensorflow/pull/9762