Replace pure-Python implementations of srgb_to_linear, linear_to_srgb,
compute_auto_levels, normalize_grayscale, dithering algorithms, and
surface_to_* conversions with calls to raygeo.image Rust functions.

The surface_to_* functions remain as thin Cairo surface buffer extraction
wrappers that delegate to raygeo.image.rgba_to_grayscale/binary/inplace.
The pure-array dither functions (apply_floyd_steinberg_dither,
apply_bayer_dither, apply_minimum_run_length) are removed from
rayforge.image.dither and imported from raygeo.image instead.

srgb.py retains create_lut_from_color and resize_linear_nd which have
no raygeo equivalents.

The Rust-accelerated image processing completes much faster than the
previous Python implementation, causing rapid-fire invalidations to
flood the pipeline with overlapping task replacements. This leaves
nodes stuck in PROCESSING state and prevents settle.

Raise min_invalidation_interval_ms from 2ms to 50ms to give the
pipeline breathing room between invalidations. Also increase
max_settle_wait_sec from 45s to 60s.

The is_busy property iterated ALL contexts (including superseded ones)
to check for active tasks. When rapid invalidations cause generations
to be superseded before their tasks complete, orphaned task counts in
old contexts kept is_busy True forever, preventing the pipeline from
settling.

Only check the active context's task counts. Superseded contexts clean
up their own tasks independently and should not block the pipeline's
busy state.

- Mark nodes as DIRTY on cancellation so scheduler re-processes them
- Fix is_busy tests to match active-context-only behavior

The 'stuck' count was just workers actively processing tasks — not
actually stuck. Only log when there are dead workers.

Workers now check the cancellation flag before running user_func,
allowing them to skip already-cancelled tasks instantly instead of
processing them to completion. This prevents cancelled tasks from
the chaos phase from blocking the queue and starving current-
generation tasks.

Also adds a running message from workers so _worker_start_time is
populated, enabling the per-worker timeout to detect and terminate
unresponsive workers after 30 seconds. Gate _emit_node_state calls
on is_generation_current() to prevent stale completions from
corrupting DAG state. Add _busy_reason() diagnostic method and
improve settle-phase logging.

The early cancel check broke test_cancel_propagates_to_worker which
expects the task function to run and check is_cancelled() itself.
The is_busy fix (has_tasks + context check) is sufficient to make
the stress test pass without needing to skip tasks in the worker.