Refactor MIR field/index access to projection places

[itsfuad]

Summary

MIR must keep storage projection separate from value access before arrays, slices, maps, and deeper addressable composite paths land.

First slice is implemented in #34:

• replaces tactical FieldAddr with ProjectField
• replaces StoreField with generic pointer Store
• adds explicit Load for pointer-backed field reads
• makes @box.field, box.field, and box.field = value share field projection logic

Problem

MIR used to mix place computation and value access:

• Field was value-producing.
• StoreField was a special write path.
• Address-of field needed a one-off FieldAddr bridge to avoid LLVM pointer-to-copy temps.

That shape does not scale to arrays, slices, maps, nested fields, deref/index chains, or future addressability rules.

Direction

Continue pivoting MIR to projection/place form:

• represent storage locations as projections, e.g. field, index, deref/local base
• make reads explicit with Load(place)
• make writes explicit with Store(place, value)
• make address-of reuse place/projection computation instead of one-off address nodes

Current field slice spelling:

• box.field through pointer -> Load(ProjectField(box, field))
• @box.field -> ProjectField(box, field)
• box.field = value -> Store(ProjectField(box, field), value)

Hard Line

Do not add future one-off address nodes:

• no IndexAddr
• no SliceAddr
• no MapAddr

Arrays, slices, and maps must extend the projection/place path instead.

Remaining Acceptance Criteria

• Add non-field projections needed by arrays/slices before indexing lands.
• Keep backend GEP/load/store logic shared instead of duplicated by composite kind.
• Make map element addressability policy explicit before map lowering lands.
• Decide whether MIR keeps projection nodes directly or grows a fuller Place model.

PRs

• First field projection slice: Add MIR field projections #34
• Raw pointer/address operator base work: Add raw pointer address escape checks #33