Taichi

- supports passing vector/matrix arguments in autodiff kernel
- supports autodiff for torch Tensor and taichi ndarray on CPU and CUDA
- supports passing grad tensor to primal kernel

3. Bug Fixes

- Improved the quality and coverage of error messages

- Added C-APIs of `ti_import_cpu_memory()` and `ti_import_cuda_memory()`
- Added support for multiple AOT runtime devices
- Added support for matrix/vector in compute graph in C-API
- Added support for matrix/vector in compute graph in Python

- Added Metal backend support for GGUI

Added helper function to create a 2D/3D sparse grid, for example:

create a 2D sparse grid
grid = ti.sparse.grid(
{
"pos": ti.math.vec2,
"mass": ti.f32,
"grid2particles": ti.types.vector(20, ti.i32),
},
shape=(10, 10),
)

access
grid[0, 0].pos = ti.math.vec2(1, 2)
grid[0, 0].mass = 1.0
grid[0, 0].grid2particles[2] = 123

Added the following `ti.simt.block` APIs:
- `ti.simt.block.sync_any_nonzero`
- `ti.simt.block.sync_all_nonzero`
- `ti.simt.block.sync_count_nonzero`