fix(analysis): make access relations state what their Ops read

[zhen8838]

Why

• Fixes fix(ir): make five access relations state what their Ops read #207
• Five registered access relations stated reads and writes their Ops do not perform: full_like read its template, ShapeOf/Rank/Local had wrong boundary ranks, RoPE accepted q/k of different ranks until isl failed, the Reduce workspace was addressed by source coordinates, and Copy rejected shard shapes its verifier accepts.
• Placing them correctly needed one IR→isl conversion API and a real layout placement (positions_of dropped merged axes and folded out-of-range coordinates onto legal positions).
• Correct numbers then exposed two existing defects: MatMul gave a replicated result its lhs layout domain, and the first-solution allocator seed overshot the live-byte bound (including a false smem capacity error in the schedule tutorial).

What

• One carrier: AccessRelation(relation, values). A handler returns a flat tuple, one per argument then one per result field; AffineAccess/BoundaryRelation/AccessRelations and the relation_of/restricted_access/boundary_maps unpacking helpers are removed.
• ir/isl_interop.py: one IslParamValues dictionary; dim_to_isl_pw_aff, shape_to_isl_set, isl_to_dim, and layout_to_isl_map (flattened layout positions, nested shards per layer, mesh coordinates as parameters). Parameters are matched by value identity and named p<number>, never off a value's name; iterating renames each Op's parameters by identity into new relations before combining them.
• projected places shard boundaries with layout_to_isl_map, clips logical coordinates first, and fixes a requested unit (device, default 0).
• readnone_relations/unread_access for metadata Ops; full_like reads nothing and writes everything (cost evaluator synced).
• RoPE rejects q/k rank mismatch at typeinfer; Reduce workspace reaches one slot per warp and verify checks its size; Copy relates same-buffer shard shapes by a row-major reshape.
• MatMul type inference builds an unsharded result layout over its own shape.
• Allocation seed: first-fit in fixed orders, each checked against every address constraint; the most compact valid seed bounds the search, and an invalid one never does.
• Extended existing tests with hand-computed guards (tile bytes, 2·M·N·K flops, live-byte peaks, independent WGMMA max(tensor, smem) time); regenerated affected goldens and tutorials.

Contract

• docs/spec/types.md §11 (new ISL interoperability section), visitor-registry.md, semantic-analysis.md: AccessRelation tuple order, fixed p<number> names, identity within one conversion context, layout placement, readnone relations.
• docs/spec/hir.md, tir.md: RoPE same-rank q/k; Reduce workspace size; Copy between shard shapes; FullLike reads only its template's type.
• docs/spec/analysis.md: a seed is an internal candidate placement (addresses and peak); one that bounds the allocator must meet every address constraint.
• Memory reports change where old numbers were wrong: whole smem tiles, result-sized MatMul buffers, lower placement peaks; KNOWN_OVER_BOUND is now empty.

Risk

• FullLike no longer aliases its template in allocation (its old proof relied on the removed read); peaks still meet the live bound.
• The allocator seed improves fixed cases, not every graph; the solver still stops at its first feasible solution.
• test_wgmma_performance...[n=16] agrees with the model at ceil(10.93 ns) vs ceil(10.09 ns) because of different clock references.