[Lang] Customized struct support

[victoriacity]

Related issue = close #2614

Adds customized struct support to Taichi fields, e.g.:

f = ti.Struct.field({'a':ti.i32, 'b':ti.f32}, shape=(8,))
# or
# f = ti.Struct.field({'a':ti.i32, 'b':ti.f32})
# ti.root.dense(ti.i, 8).place(f)
@ti.kernel
def run():
  for i in f:
    f[i].a = i
    f[i].b = f[i].a

The implementation generally follows the Matrix and MatrixField class with similar API definitions, while members of a Struct can be only accessed using attribute names.

A CompoundType interface has been implemented to instantiate vector, matrices, and structs with pre-defined shapes or members. This allows for a custom struct to have vectors, matrices and nested structs as members.

vec3f = ti.types.vector(3, ti.f32)
ray3f = ti.types.struct(origin=vec3f, raydir=vec3f, raylen=ti.f32)
pixelrays = ray3f.field(shape=(512, 512))

Local structs in both python and Taichi scopes can be initialized with either keyword arguments and dictionaries:

ray = ti.Struct(
    origin=ti.Vector([0.0, 0.0, 0.0]), raydir=ti.Vector([0.0, 0.0, 1.0]), raylen=1.0,
)
# or
ray = ti.Struct({
    "origin": ti.Vector([0.0, 0.0, 0.0]),
    "raydir": ti.Vector([0.0, 0.0, 1.0]), 
    "raylen": 1.0,
})

In addition to directly invoking initializers of Struct and Matrix classes, CompoundType instances can be also called to create matrix or struct instances. Vectors and matrices can be created using GLSL-like broadcast syntax since the shape of the vector or matrix is already known:

vec2f = ti.types.vector(2, ti.f32)
...
ro = vec3(0.0) # ti.Vector([0.0, 0.0, 0.0])
rd = vec3(vec2(0), 1) # ti.Vector([0.0, 0.0, 1.0]), will perform implicit cast
ray = ray3f(origin=ro, raydir=rd, raylen=1.0)

[k-ye]

Ping :-) ?

[victoriacity]

/format

[victoriacity]

/format

[strongoier]

Excellent work!!

[victoriacity]

/format

[strongoier]

Would you also resolve conflicts with master?

[victoriacity]

/format

[strongoier]

LGTM! Thanks for your effort!