[ir] MatrixField refactor 1/n: Make a GlobalVariableExpression solely represent a field

[strongoier]

Related issue = #5959

Previously, a GlobalVariableExpression serves two purposes:

• representing a field;
• being a wrapper of a (maybe intermediate) SNode so that no matter a FrontendForStmt is actually a struct for on a SNode or an external tensor, it can always be constructed by passing an Expr.

However, the second purpose doesn't make sense - a struct for on a SNode and a struct for on an external tensor are handled differently everywhere (lower_ast, ir_printer, ...), so it's natural to split the constructors as well.

This PR clearly and consistently distinguishes different types of FrontendForStmt:

• a struct for on a SNode (snode != nullptr);
• a struct for on an external tensor (external_tensor != nullptr);
• a mesh for (mesh != nullptr);
• a range for (other).

As a result, the second purpose is eliminated and a GlobalVariableExpression serves solely for a field.

[ailzhang]

LGTM!
nit: external_tensor isn't the best esp. when it comes to ndarrays. But this naming convention is kinda misleading but consistent in the current system lol. Shall we consider renaming them all together in a separate PR in the future?

[strongoier]

LGTM! nit: external_tensor isn't the best esp. when it comes to ndarrays. But this naming convention is kinda misleading but consistent in the current system lol. Shall we consider renaming them all together in a separate PR in the future?

Sounds good. There are other misleading names like GlobalPtr, and we might need to systematically rename them all.