refactor(ir): add AtomSched operation declarations

[zhen8838]

Why

• Establish the Stage 2a IR declarations and package boundaries required before migrating AtomSched operations in Stage 2b.
• Make mesh composition, scheduling patterns, memory effects, pointer views, and dynamic tuple selection explicit contracts instead of downstream overrides.

What

• Add layered mesh composition, slicing, separation, and topology validation.
• Move the pattern language into ir.pattern, rename constraints to clauses, and add match/constraint utilities plus mesh and layout patterns.
• Add ParamDef.effect, PointerType, pointer-based memory operation declarations, and Copy/Fill/Cast scheduling metadata.
• Make TupleGetItem.index an input and support dynamic tuple indexing through parsing, printing, evaluation, analysis, and CUDA emission.
• Update specifications, fixtures, tutorials, and all affected call sites.

Contract

• make_mesh now composes Mesh values; separate reverses layered composition, and multi-level meshes support per-level slicing.
• tilefoundry.ir.core.pattern is removed in favor of tilefoundry.ir.pattern; DimVarRangePat is now RangePattern, including the public DSL export.
• tilefoundry.ir.constraints is renamed to tilefoundry.ir.clause with the corresponding public type and metadata names.
• Op inputs may declare ParamDef.effect; PointerType is part of the public Type union.
• T.tensor_view consumes a pointer operand. Pointer-view trailing coordinates are explicitly unsupported.
• TupleGetItem.index is a scalar input. Dynamic indices require a non-empty homogeneous tuple and conservatively read all tuple leaves.

Risk

• Dynamic tuple CUDA emission uses C++20 aggregate CTAD for cute::array{...}; Stage 2b must validate it in a compiled dynamic-stage fixture.
• Dimension arithmetic rendering remains duplicated between tuple indexing and tensor-view coordinates; Stage 2b should consolidate the two tables when it next changes tensor-view codegen.