Lands a self-contained cine pipeline for single-file ultrasound DICOMs
(SOP Class UID 1.2.840.10008.5.1.4.1.1.3 / .3.1, NumberOfFrames > 1)
alongside the existing volume pipeline. Multi-chunk volume imports never
match the cine router, so CT/MR streaming and 3D volume behavior are
unchanged.

Core additions under src/core/cine/:
- parseCineDicom.ts wraps cornerstonejs/dicom-parser to extract the
  header (transfer syntax, geometry, FrameTime, patient/study/series,
  SequenceOfUltrasoundRegions) and per-frame byte views — zero-copy for
  native PixelData, fragment-aware for encapsulated JPEG-Baseline (with
  populated BOT, empty BOT, and JPEG-SOI scan fallbacks). Supports
  Implicit + Explicit VR LE and JPEG-Baseline.
- DicomCineImage extends BaseProgressiveImage, owns one 2D vtkImageData
  (extent [0,cols-1, 0,rows-1, 0,0], 3-component RGB uint8), and swaps
  scalars in-place when the selected frame changes. setFrame() bumps a
  decode token unconditionally so any new request — cached or decode —
  invalidates in-flight decodes.
- frameCache.ts: byte-budgeted LRU keyed by frame index; decodeJpegFrame
  via createImageBitmap + OffscreenCanvas; decodeNativeFrame for native
  RGB / MONOCHROME2.
- isCineImage / getCineImage helpers and getRenderSlice that returns 0
  for cine so the VTK mapper renders slice 0 while the semantic slice is
  the frame index.

Import routing in src/store/datasets-dicom.ts: when a chunk group has a
single chunk, an UltrasoundMultiframe SOP UID (current or retired), and
NumberOfFrames > 1, it diverts to _importCineChunk before the legacy
DicomChunkImage path. View integration:
- VtkBaseSliceRepresentation.vue: render-slice helper pins VTK slice to
  0 for cine; conditional W/L sync (cine pixels are display-encoded, so
  wlConfig defaults don't clobber them); slice watch is immediate so
  restored sessions paint the saved frame on first mount.
- SliceViewerOverlay shows "Frame: N/M" and hosts a new CineTransport
  component (play/pause/loop/fps via useIntervalFn, FPS seeded from
  FrameTime).
- Ruler/Rectangle/Polygon widgets use getRenderSlice for their plane
  manipulator origin so annotations still scope to a frame.
- view-configs/slicing.ts overrides the slice range to
  [0, numberOfFrames - 1] for cine.
- image-cache.removeImage now calls dispose() before delete; cine's
  dispose clears the LRU, drops compressed frame refs, and deletes the
  vtkImageData.
- image-stats early-returns for cine ids — histogram/auto-range is
  meaningless on display-encoded data.

Testing:
- 3 vitest tests build a synthetic DICOM in-memory (Explicit VR LE
  native, encapsulated with populated BOT, encapsulated with empty BOT)
  to exercise the parser without external fixtures.
- New tests/specs/cine-rendering.e2e.ts loads US-MONO2-8-8x-execho.dcm
  from the BSD-licensed GDCM corpus on SourceForge (cached via the
  existing wdio onPrepare hook into .tmp/), asserts the cine transport
  renders with "1 / 8", and asserts the counter advances on ArrowUp.

Adds dicom-parser ^1.8.21 (MIT, 0 deps, ~6.9 KB gzipped) to
devDependencies.