kbuild: fix config make targets being merged as fragment data

[nuclearcat]

The kselftest build fragment is a single entry, make:kselftest-merge
(kernelci-pipeline config/fragments.yaml), which tuxmake accepts as a magic
--kconfig-add value (kernelci/tuxmake#275, documented since kernelci/tuxmake#280).

_parse_fragments() writes every fragment entry into fragments/N.config and
passes that file to --kconfig-add, so the directive reaches kconfig as
fragment content and gets merged as unexpected data:

.config:5652:warning: unexpected data: make:kselftest-merge

No kselftest config has been merged since the fragment was switched over, on
either backend. The resulting kernels are missing everything
tools/testing/selftests/*/config asks for — CONFIG_IOMMUFD,
CONFIG_FAULT_INJECTION, CONFIG_USER_NS, CONFIG_BPF_SYSCALL,
CONFIG_VETH, ... — which wipes out whole suites rather than individual tests.

Observed effect in production: kselftest-iommu has 1 pass in 18376 runs
all-time, and accounted for ~43k of the ~61k test-level failures in the
04-07 Sep window (70%), pushing the overall test failure rate from ~3.2% to
9.8%. Every subtest dies in fixture setup on a missing /dev/iommu:

# # iommufd.c:66:simple_close:Expected -1 (-1) != self->fd (-1)

with # CONFIG_IOMMUFD is not set in the build's .config, on hardware whose
IOMMU is otherwise fine (DMAR: dmar0 ..., iommu: Default domain type: Translated).

Changes

1. kbuild: pass config make targets through to tuxmake — split make:
   directives out of fragment content and keep them as separate kconfig
   additions, ordered with the fragment files so merge order is preserved, then
   pass them to tuxmake verbatim. _fragment_files becomes _kconfig_adds
   since it no longer holds only files. The make backend skips them, unchanged,
   so this commit stands on its own.
2. kbuild: run config make targets in the make backend — emit
   make <target> in place of the merge_config.sh call. kselftest-merge
   merges into the .config built so far and runs its own olddefconfig, so it
   drops into the existing sequence.

Result

tuxmake: --kconfig-add=.../0.config --kconfig-add=.../1.config \
         --kconfig-add=make:kselftest-merge --kconfig-add=.../3.config

make:    make x86_64_defconfig
         merge_config.sh -m .config .../0.config
         merge_config.sh -m .config .../1.config
         make kselftest-merge
         merge_config.sh -m .config .../3.config
         make olddefconfig

Testing

ruff check and ruff format clean; tests/test_kbuild.py passes (13 tests,
4 new in TestFragments: directive not written to a fragment file, mixed
config/directive fragments split correctly, tuxmake passthrough, make-backend
ordering).

No kernelci-pipeline change is needed — config/fragments.yaml was already
correct; only the consumer was dropping the directive.

Draft: worth a build on staging to confirm CONFIG_IOMMUFD lands in
x86_64_defconfig+lab-setup+x86-board+kselftest before merging, and to check
that the now-effective kselftest-merge doesn't pull in config that upsets
other suites (it is all-or-nothing across every subsystem).