[Feature][IR][Dataflow] Keep `meta_for` in IR

[Fangtangtang]

Motivation and Expected Behavior

Currently, meta_for loops are fully unrolled when building IR. We want to preserve meta_for loop structure at the highest-level IR.

This helps:

• simplify builder for meta programming (move the burden to 'lowering' pass)
• allow backends decide if and when meta_for loops should be unrolled. (e.g., AIE may want to keep loops rolled to reduce code size [IR][AIE] Support rolled meta_for loops to reduce code size #429)
• aligns with the plan to introduce an operation for meta_if ([Feature][IR][Dataflow] Keep meta_if / meta_else in IR to preserve 'singel df.kernel' program structure #523)

We plan to introduce a new operation allo.meta_for (following the design of scf.for).

Related resources

No response

[Fangtangtang]

We can directly use affine.for in high-level IR, and use unrolling pass to fully unroll the meta for loops. (some existing pass like -affine-loop-unroll can be useful).

This helps simplify our dialect and IR