0.5.3 — Round-4 parity fixes

A regression release. A workspace-wide upstream-parity audit (silx / PyDM /
adl2pydm + MEDM C) found that three of the fixes 0.5.2 shipped were wrong, and
turned up ten further divergences from the reference behaviour. All are fixed
here. VolumeRaycaster users should upgrade: the volume it rendered in 0.5.2
carried no thickness cue and lost the hue of faint voxels.

Fixed — rsplot

• Volume opacity depends on the thickness a ray crosses again. 0.5.2's
  "Beer-Lambert" correction divided by the step count, an exponent carrying no
  world distance, so every ray accumulated the same transmittance whatever chord
  it traversed: a uniform volume rendered as a flat silhouette, and at the
  default 256 steps the correction was exactly the identity. Each sample is now
  corrected from a reference spacing (box_diagonal / 256, a world distance),
  which keeps the step-count invariance 0.5.2 was after while restoring the
  density cue. A ray crossing the full box diagonal at the default step count
  renders exactly as before.
• Faint voxels keep their colour. The premultiplied 3-D texture 0.5.2
  introduced stored the product rgb · a in 8 bits, and the shader must divide
  the coverage back out to recover the straight colour — a product that at
  coverage a resolves the quotient only to steps of 1/a. An authored
  (255, 128, 0) voxel at a = 3/255 read back as (255, 170, 0); at
  a = 1/255 no hue survived at all. The texture is now Rgba16Float, which
  carries the straight colour exactly at every coverage while the filter still
  interpolates premultiplied colour (no return of the dark fringe). This
  doubles the volume texture's VRAM.
• VolumeRaycaster::remove no longer blanks another view. Two views sharing
  a VolumeId share one texture, by design; remove freed it outright, so the
  survivor silently rendered nothing. The texture is now claim-counted and lives
  exactly as long as the views holding its id. Dropping a view releases its
  claim too, so a VolumeRaycaster that is simply dropped no longer pins its
  VRAM for the app's lifetime.
• Alt+Wheel zooms X only and Shift+Wheel zooms Y/Y2 only, the silx bindings
  the axes menu already advertised. Plain wheel and keep-aspect are unchanged.
• Middle-button drag pans in every interaction mode, as in silx. The
  existing right-button pan is kept.
• The Reset-Zoom toolbar button disables when neither axis autoscales, and
  its tooltip names the single autoscaling axis when exactly one is on.
• Programmatic set_limits and toolbar Zoom-In/Out repair their limits.
  Inverted, degenerate or float32-overflowing ranges reached the transform
  unrepaired; every view-limits commit now runs the same clamp, including the
  Y2 and extra-axis paths.
• The colormap dialog's histogram drops samples equal to the range maximum,
  as Histogramnd does with its default half-open last bin.
• Scatter BINNED_STATISTIC drops points on the max edge instead of
  clamping them into the last bin, matching scipy.stats.binned_statistic_2d
  as silx calls it. A degenerate extent now admits no point rather than
  collapsing every point into bin 0.

Fixed — rsdm

• calc:// re-evaluates when a child connects or disconnects. PyDM
  re-emits on a connection-state change; rsdm only watched values.
• calc:// no longer retriggers on a child's alarm or metadata change, only
  on an actual value change, as PyDM's update list specifies.
• A dialect = medm expression reading operand I or J re-evaluates on an
  alarm status / severity change. MEDM's setDynamicAttrMonitorFlags adds
  monitorStatusChanged and monitorSeverityChanged exactly when the
  expression uses those operands. rsdm never watched severity, so a transition
  left the calc showing a stale value, and it had no alarm status to read at all
  — operand I bound a constant 0.0. Both are now monitored, and only when
  the expression reads them.

Added — rsdm

• ChannelState::status carries the wire alarm status — the record's STAT
  for ca://, the pvAccess alarm.status enum for pva://. It backs the MEDM
  calc operand I.

Fixed — adl2rsdm

• Embedded-display and composite-file children take the host screen's
  colormap. MEDM's parseCompositeFile parses and discards the child file's
  "color map" block and resolves the children's clr/bclr against the
  parent display's palette; adl2rsdm used the child's own. Screens on the
  default palette are unaffected; those with a custom palette rendered every
  embedded widget in the wrong colours.
• An arc with no path spans 90°, not 360° — MEDM's createDlArc default.

Changed — rsplot

• The volume 3-D texture is Rgba16Float rather than Rgba8Unorm, doubling its
  VRAM. See the faint-voxel fix above for why an 8-bit texture cannot carry the
  data.
• VolumeRaycaster now holds a clone of its RenderState (a bundle of Arcs)
  so it can release its texture claim on drop.