Refactor check_light_mesh_visibility for performance

[re0312]

Objective

• The current check_light_mesh_visibility system has unsatisfactory performance, as it involves multiple hash lookups for every mesh entity and lacks parallelization

Solution

• split check_light_mesh_visibility to check_dir_light_mesh_visibility and check_point_light_mesh_visibility

• check_dir_light_mesh_visibility defers setting the entity's viewVisibility so that Bevy can schedule it to run in parallel with check_point_light_mesh_visibility.

• Reduce HashMap lookups for directional light checking as much as possible

• Use par_iter to parallelize the checking process within each system.

Testing

cargo run --release --example many_cubes --features bevy/trace_tracy -- --shadows

command apply:

reduce from 4.88ms to 972+ 303 =1.2ms ,~4x speed up.

cargo run --release --example many_foxes --features bevy/trace_tracy -- --count 10000

command apply:

reduce from 646us to 142+ 97 =239us ,~3x speed up.

In the most common scenarios, there is usually one directional light and multiple point lights, and this PR should provide more advantages as direction light checking can run in parallel with point light checking.

[IceSentry]

Could you maybe make a separate PR for the split? It's a bit hard to review the other changes because it makes the diff very noisy

[re0312]

Could you maybe make a separate PR for the split? It's a bit hard to review the other changes because it makes the diff very noisy

make sense, i split this pr to #13905 and #13906 for better review.

[alice-i-cecile]

This has now been split :)