Support type specification and polymorphism in internal functions

[re-xyr]

It will be better if we are able to implement more operations (unary/binary ops) as internal functions instead of hard-coded operations. Because these operations are polymorphic on argument types (size & int/float) and perform implicit casts, we need to extend the functionalities of internal functions to allow type specification and polymorphism.

(cc @strongoier)

[re-xyr]

Some points I gathered till now:

• Eventually we'd like to support Unary/Binary/TernaryOps with funccalls.
  • Unary and binary operators all have multiple type signatures.
  • Moreover, ternary op's return type is 1) polymorphic and 2) can only be primitive.
  • Both of these suggests that we need polymorphism and also something like "traits" that can be attached to type variables, like Primitive, Integral, Floating, etc.
  • All of these are under the assumption that Taichi will eventually scrap implicit conversion.
• ExternalTensorShapeAlongAxis is a very tricky one. Its typechecking is not checking types.
  • Instead it requires the operand to be a specific Expr construct, namely ExternalTensorExpression, which... seems not to have a type.
  • We need to 1) give ExternalTensorExpression a type and 2) have an assertion that ExternalTensorExpression will be the only construct to ever have this type.
  • Otherwise we could also just make the parameter to be the arg_id of the tensor instead of the tensor itself.

In all the current taichi typechecking is a quite atypical one. Some expressions (especially, non-primitive ones) don't even have types and maybe we should fix that first.

[re-xyr]

Hot take: If we're unable to give an Expr a type, it probably should not be an Expr at all, but some separate construct