Modern DCMTK (>= 3.6.0) ships DCMTKConfig.cmake and DCMTKTargets.cmake,
exporting DCMTK::* imported targets with INTERFACE_INCLUDE_DIRECTORIES
and INTERFACE_LINK_LIBRARIES already wired. The hand-rolled module-mode
FindDCMTK.cmake predates that and only worked against installs that
exposed the legacy DCMTK_INCLUDE_DIR / DCMTK_LIBRARIES variables.

* Delete config/cmake/Modules/FindDCMTK.cmake.
* Replace the include() of that module in core/vil/CMakeLists.txt with
  find_package(DCMTK CONFIG REQUIRED).
* Link vil against DCMTK::dcmimgle, DCMTK::dcmdata, DCMTK::ofstd
  instead of the flat DCMTK_LIBRARIES list; include dirs propagate via
  the imported targets.

Consumers point DCMTK_DIR at an installed DCMTKConfig.cmake (or place
DCMTK on CMAKE_PREFIX_PATH) as with any other modern CMake package.

, vxl#769, vxl#816)

vil_dicom has not built against DCMTK >= 3.6.0 for roughly a decade
because of structural API changes upstream. With the bundled v3p/dcmtk
now gone, vil_dicom must speak the modern DCMTK API. This commit
delivers a port that builds and runs cleanly against DCMTK 3.7.0.

API drift addressed:

* Namespaced headers. DCMTK 3.6+ ships its headers under
  dcmtk/<library>/<header>.h. All references in vil_dicom.cxx are
  updated to the namespaced form.

* DiInputPixelTemplate signature change (vxl#550). The constructor went
  from 6 to 9 parameters and now reads pixel data through a DiDocument
  rather than a raw DcmPixelData*. The reader constructs a DiDocument
  wrapping the already-loaded DcmFileFormat at the read site and
  threads it through read_pixels_into_buffer and convert_src_type.
  vil only loads a single 2D slice, so frame_count=1,
  frame_size=num_samples, with a stack DcmFileCache and a zero
  fragment index.

Input-stream wiring:

The historical approach -- subclass DcmInputStream / DcmProducer /
DcmInputStreamFactory inside vil to adapt a vil_stream -- compiles
cleanly against DCMTK 3.6+ once eos()/avail()/read()/skip()/putback()
adopt offile_off_t and DcmInputStreamFactory::ident() is implemented
(vxl#769). It segfaults at runtime, however: DCMTK's static libraries
are built with -fvisibility=hidden plus a private set of POSIX
feature-test macros (_DARWIN_C_SOURCE, _XOPEN_SOURCE_EXTENDED,
_BSD_SOURCE, _BSD_COMPAT, _OSF_SOURCE) that DCMTK does not propagate
via INTERFACE_COMPILE_DEFINITIONS on its imported targets. A consumer
that derives from DcmInputStream therefore sees a slightly different
class layout than the one frozen in libdcmdata.a, corrupting the
input-stream vtable on the first virtual dispatch from
DcmFileFormat::readUntilTag. Mirroring DCMTK's internal flags from
the consumer side is fragile and platform-specific.

The robust fix is to keep every DCMTK-derived object inside
libdcmdata.a: spool the vil_stream buffer to a mkstemp() temp file,
hand the path to DcmFileFormat::loadFile(), and unlink the temp file
immediately. loadFile() uses DCMTK's own DcmInputFileStream entirely
inside the library, so no DcmInputStream-derived types remain on the
consumer side and the ABI hazard is eliminated. The vil_dicom_stream
{factory,producer,input} classes and their header are dropped.

Build verified against DCMTK 3.7.0 on macOS arm64 Apple Clang
Release -O3 -std=c++17. core/vil's full test suite (84 cases in
test_file_format_read including 6 DICOM cases) passes; see the
follow-up commit that adds a focused vil_dicom smoke test.

test_dicom drives vil_dicom_file_format::make_input_image against the
988-byte 16-bit DICOM fixture that has been shipped in
core/vil/tests/file_read_data/ for years but never exercised, since
no vxl build was wired against DCMTK until now. It complements the
DICOM cases already in test_file_format_read by exercising the
reader directly (bypassing the full vil_load multi-format probe),
giving a focused regression signal if the DCMTK port drifts again.

The new test_dicom is registered under CTest as vil_test_dicom only
when VXL_USE_DCMTK is ON. When DCMTK support is off, the test source
compiles to a stub that prints a skip message, so the test driver
keeps a stable shape across configurations.

The error branch invoked dset.func(key, var) a second time solely to
print its OFCondition text, repeating the dataset lookup and double-
evaluating macro arguments with side effects. Capture the condition
once and reuse it.

A truncated or corrupt encapsulated pixel data fragment can leave
loadAllDataIntoMemory() returning non-EC_Normal while loadFile()
itself succeeded. The previous code dropped that condition and
unlinked the temp file regardless, so downstream decode would later
try to reopen the deleted path. Fold the second condition into the
same OFCondition that already gates the error branch below.

When dset.search(DCM_PixelData,...) returned non-EC_Normal the
constructor left 'pixels' at its initial null value and passed it to
read_pixels_into_buffer, which dereferences pixels->getVR() before any
null check, causing a SIGSEGV on datasets such as DICOMDIR, structured
reports, and RT-struct files.

Per-image registerCodecs()/cleanup() churns process-global DCMTK
state and can race with or tear down codec registrations held by
co-resident DCMTK consumers (ITK, GDCM-adjacent components). The
paired bool flags also leaked the registration on any early return
between register and cleanup.

Replace the conditional per-image register/cleanup with a single
function-local static initializer that registers dcmjpeg, dcmjpls,
and dcmrle exactly once per process. RLE Lossless (EXS_RLELossless)
now decodes correctly; previously usesEncapsulatedFormat() steered
RLE into the baseline-JPEG branch, registering a codec that cannot
decode RLE and silently producing a null image. Registration is
idempotent and thread-safe via C++11 magic statics, so no per-image
cleanup is needed.