Add manual seam positioning + a seam debug line

[RufanMelfor]

Reimplemented from scratch against the current architecture (post
"stitch unification" / "executor spine" refactors) — the original
implementation predates both, and neither seam_offset nor a debug-line
concept exist anywhere in the current codebase.

Core addition: Topology::seam_offset (f32, default 0.0, bounded to
±SEAM_OFFSET_RANGE = 0.3), following the codebase's established pattern
of homing render-affecting rig state on the Calibration document
(alongside blend_width) rather than on ViewportConfig. It shifts only
the alpha-crossfade threshold within the coverage the calibration already
guarantees — never the plane geometry itself, so it can't open a black
gap the way moving intersect could.

Threaded through both render paths that must agree by construction:

• GPU: fisheye.wgsl's previously-unused lens_preview.z slot now
  carries seam_offset; lens_preview.w carries show_seam_line (a
  debug line drawn by reusing the exact same edge_dist the alpha blend
  computes, so it can never disagree with where the blend actually sits).
• CPU: BlendRule::Smoothstep gained an offset field (was a bare
  width) — this is the software stitcher that must mirror the shader
  exactly for the agreement-oracle tests. Added
  cpu_and_gpu_backends_agree_with_seam_offset (a real end-to-end GPU
  test on real hardware, not just a compile check) alongside a focused
  BlendRule::alpha unit test proving the offset shifts the curve without
  changing its shape.

show_seam_line is deliberately not calibration state (lives on
StitchPipeline directly, GPU-executor-only — the CPU executor is a
headless correctness oracle with no interactive preview to draw a debug
overlay onto).

Wiring: StitchCore::set_seam_offset/set_show_seam_line →
StitchSession → StitchJob::seam_offset()/show_seam_line() builders →
reco stitch --seam-offset/--show-seam-line → reco-gui "Seam position"
slider + "Show seam line" checkbox, mirroring the existing "Seam blend"
slider's wiring exactly (including calibration-load/reset sync).

Verified: fmt --check clean; cargo test -p reco-core (153 passed,
including both new tests; only the 2 pre-existing CUDA hardware-gap
failures, unrelated); full workspace build clean.

[github-actions]

All contributors have signed the CLA. Thank you!
_{Posted by the CLA Assistant Lite bot.}

[RufanMelfor]

I have read the CLA Document and I hereby sign the CLA