fix: use camera-relative coordinates to avoid float precision loss at large world coordinates

[Whisperlyric]

Problem

At large world coordinates (tens of millions, e.g. near the world border):

• the trajectory line breaks up into discrete, disconnected segments, and
• the impact block outline is distorted — it no longer forms a 1×1×1 cube
  (some edges stretched to several blocks), so the
  highlighted target can no longer be identified.

Cause: all render vertices were converted to float in absolute world
coordinates (new Vector3f((float) pos.x, ...) for trajectory points,
float-cast box corners in RenderUtils). A float only has a 24-bit mantissa,
so at ~30,000,000 the spacing between representable values grows to 2–4 blocks:
adjacent line vertices snap onto the same coarse grid, which shatters the line,
and a box's min/max corners are rounded independently, so its edges no longer
span exactly one block.

Solution

Subtract the camera position on the CPU in double and convert the resulting
camera-relative offsets to float instead; relative coordinates are small, so no precision is lost. The now-redundant poseStack.translate(-cam) is removed. Applied to: - trajectory line segments and the impact cross (PtpClient.renderTrajectory) - highlight/outline boxes (RenderUtils.renderFilledBox / renderBox)