Support data types other than f32

[pixelspark]

This PR (based on my earlier PR #45 that implemented the DAG) adds support for data types other than f32. Right now I have implemented support for i32 inputs, but others (such as f16) should be fairly easy to add. The current version will always cast outputs to Vec<f32> but of course the output type should depend on the actual output type in the model.

Notably, BERT and some other models appear to use int64 inputs - working with i64 in WGSL however is not (yet?) possible. We should therefore probably re-interpret these as int32.

Some implementation notes:

• Shape now has a data_type field of type ScalarType. The latter holds some information about the type's size and striding.
• In struct.wgsl, I define a type Scalar to be the 'default' scalar type (all ops except those that do conversions work in a single type) based on template variable scalar_type. It also defines Vec3, Vec4, Mat4x4 etc. which are the appropriate vector/matrix types. All shaders use these type names instead of the raw ones.
• The compiler sets scalar_type variable by looking at the types of inputs/outputs. They should all agree (except in cases where there may be inputs/outputs of different type, which happens e.g. for Reshape where the first input is the data and the second input is a shape, with integer data type) or else an error is thrown.
• Input tensors are now wrapped in an enum, e.g. InputTensor::F32(my_input.as_slice()). Output tensors are currently still all of type Vec<f32> (when the actual output data type is different the current version casts the elements to f32 first).

See tests/arithmetic.rs for an example of integer addition.

[pixelspark]

@haixuanTao I rebased this on top of your workspace PR, this one should be ready to go now.

[haixuanTao]

Ok I think we can merge this branch.

[pixelspark]

Ok I think we can merge this branch.

Yes, agreed. We might improve the API in a later PR (e.g. the builder pattern or something else)