fix(macos): prevent bidirectional capture loops

[mrcha033]

Problem

On a bidirectional Linux/Wayland ↔ macOS setup, Lan Mouse's macOS event tap
observes the pointer events posted by its own emulation backend. A synthesized
edge event can therefore be mistaken for a new local capture and immediately
send the pointer back, producing an Arch → macOS → Arch capture loop.

Changes

• tag every macOS event posted by the emulation backend and retain a process-ID
  fallback when CoreGraphics does not preserve the tag
• keep separate physical and emulated edge-arm state
• ignore the initial outward crossing, then re-arm on the first inward motion
  so short reversals work
• send a synthesized return crossing only to the matching EnterOnly capture
  while preserving normal default capture for chained topologies
• handle dragged-pointer crossings as motion
• make capture state updates synchronous and release the state lock before
  bounded event delivery
• add routing, short-reversal, drag, and stream-progress regression tests

Validation

• cargo fmt --all -- --check
• cargo test --workspace --all-features --locked
• cargo clippy -p input-capture --all-targets --all-features --locked -- -D warnings
• macOS app bundle launches at commit 3d202975
• recursive @rpath dependency closure and codesign --verify --deep --strict
  pass locally

Physical Arch → macOS → Arch verification is still running against the Arch
binary from #471 (83e66bc), so this is a draft.

Scope

• the unsuccessful layer-shell re-entry experiment is not included
• macOS dylib packaging changes are not included
• this patch targets one active incoming peer; carrying peer identity through
  CoreGraphics event provenance is left for a follow-up

This is independent of #471, which contains the input post-processing changes.