[RZE-16] feat: add layered compositor to render loop

[FrkAk]

Summary

Implements [RZE-16] — replaces the single-effect-per-handle render path with a layered compositor that alpha-blends up to MAX_LAYERS=4 effect instances per DeviceHandle into the pre-allocated canvas, then writes to hardware via the existing write_frame.

• New module src/razer_effect/compositor.py houses Layer, LayerStack, blend, and build_layer_stack. LayerStack.append raises ValueError once the cap is reached; blend does src-over alpha compositing bottom-to-top using only pre-allocated buffers (stack._blend_tmp + each Layer.scratch).
• src/razer_effect/main.py now reads cfg['profiles'][cfg['active_profile']]['layers'] instead of the v1 flat cfg['effect']. Structural reload detection is keyed off a (profile, ((effect, opacity, enabled), ...)) signature so opacity/enable/effect changes rebuild stacks while param-only edits route through effect.configure. Rescan reconciliation keeps surviving handles' stacks alive.
• Backward compat (AC4) is delivered entirely by RZE-19's _migrate_v1_to_v2 — load_config always hands main.py a v2 shape, so no shim lives in main.py or compositor.py.
• Stacked on composer/RZE-18/attempt-1 (per-handle dispatch) which is stacked on composer/RZE-19/attempt-1 (v2 config schema). PR base is set to RZE-18.

Type of change

• New feature / effect
• Refactor / cleanup

Testing

• Tested on real hardware (device model: _____) — AC1 visual smoke deferred; hardware unavailable in this environment.
• OpenRazer version: _____
• Daemon running without errors after change (logic-level)
• GUI reflects changes correctly (if applicable) — layer editor lands in RZE-39.

AC verification via uv run python -c … snippets (no test suite yet — RZE-14 scope):

• AC1 logic: two static_color layers (red @1.0, blue @0.5) blend to [127.5, 0, 127.5] across every pixel; tracemalloc shows ~1.8 KB delta over 120 ticks (tracemalloc bookkeeping itself, no per-tick numpy allocations). Mixed static_color + wave stack composites with ~1.5 KB delta over 120 ticks.
• AC2: LayerStack(6, 22) accepts 4 layers; the 5th append raises ValueError(\"LayerStack cannot exceed MAX_LAYERS=4 layers\").
• AC3: top-layer opacity 0.0 yields the bottom layer; 1.0 yields the top layer; 0.5 yields the per-channel midpoint.
• AC4: v1 flat config ({effect: 'static_color', hue: 240, ...}) round-trips through load_config and produces a single-layer stack with effect_name='static_color', opacity 1.0, enabled True.

uv run ruff format . + uv run ruff check . clean. uv run ty check reports only the pre-existing openrazer.client unresolved import in device.py (unrelated).

Notes for reviewer

• The new compositor.py module is the single source of truth for blend math; future effects (ripple, breathing, twinkle, gradient_sweep, ambient, system_status, gaming_highlight) plug into it unchanged.
• Depends on RZE-18 PR feat: extend multi-device canvas with per-handle dispatch #3 + RZE-19 PR feat: extend config schema for v2 fields #2 (stacked). Merge order: RZE-19 → RZE-18 → RZE-16.
• The non---loop one-shot path now reads the active profile's first layer (falling back to key_shuffle) instead of unconditionally instantiating KeyShuffle.
• Empty-stack handles skip write_frame (matches v1 semantics of leaving unsupported devices untouched). Configurable later if we ever want a "blank on no-effect" mode.
• Tracemalloc delta over 120 ticks is well under one 64-byte numpy allocation; the blend path uses only np.copyto, np.multiply(..., out=...), and np.add(..., out=...).