release: 1.4.0 - XDMF parser, annotation mode, faster playback

[sssilvar]

Summary

Release 1.4.0. Adds XDMF support for dynamic simulation data, reworks the playback bar, and ships the previously unreleased annotation mode.

XDMF parser (.xdmf / .xmf)

• XML light data via vtkXMLUtilities; heavy data via the HDF5 C API using the copy VTK already links (VTK::hdf5, falling back to find_package(HDF5)) - no new dependency.
• Static Topology/Geometry read once (Triangle/Quadrilateral/Line, XY or XYZ); the temporal collection becomes one dataset path per step per Attribute, Node- and Cell-centered.
• Unsupported topologies, geometry types and non-HDF DataItem formats fail with an explicit message.

Playback

• Per frame: one H5Dread directly into the existing vtkFloatArray buffer (float64 to float32 on the fly), no allocation, no geometry re-read; only the array is marked modified.
• Frames cache in RAM as float under a 512 MB budget, so looping a typical run is zero-I/O.
• Playback is now clock-driven (QElapsedTimer + anchor step): each tick jumps to the frame the elapsed time calls for, dropping frames instead of queueing them, so 8x/12x runs at the requested rate rather than in slow motion.
• Bar gains previous/next frame buttons and a speed drop-up (1x/2x/4x/8x/12x, opening upwards since the bar is bottom-centered). Icons stay vector-drawn QPainterPath glyphs.
• TemporalSource is now an interface with VTKHDFTemporalSource and XdmfTemporalSource backends; parsers expose their series through MeshParser::temporal(), replacing the dynamic_cast in MeshLoading.

Verification

Loaded a 302-step FEniCS output (22801 points, 45000 triangles): geometry, scalars and playback all correct, and the fixed color range matches an independent h5py computation exactly (-84.28 .. 35.02).

Also included: annotation mode (--annotate) and mesh JSON surfaces key, both previously committed but unreleased.