feat(analysis): add memory footprints and reuse windows

[zhen8838]

Why

• Fixes fix(analysis): count an invariant operand's re-reads in traffic total #168.
• Call and Function memory records had reserved footprint slots but no unique-address working-set result, and a one-iteration L2 snapshot could not answer whether reused data stays resident.

What

• Compute Call and Function footprints by unioning final-source addresses over one target wave, preserving view resolution, packed widths, boundary movement, and completeness.
• Derive time and space reuse axes from access relations, then report each buffer's residency window, all-buffer holds, reused bytes, and the cache-capacity fits conclusion.
• Report the program-dependent wave in the header and format every text byte count through one shared unit ladder.
• Make MeshRegion an IterationScope owner, so a Call under a mesh finds its own mesh without a second IR walk, and route reuse through one memoized boundary record per access instead of replaying each relation five times.
• At realistic size, Gemm_MNK_NN128x128x64_w12x11_k16384 reports b holds=176.00MB fits=no against a 47.68MB L2, a concrete capacity finding rather than a toy-only result.

Contract

• Reorganize analysis spec 1.2.2 by record object, and specify the footprint and reuse-window conclusions together with which relation view answers each one: device-wave versus unit view for the space axis, raw narrow for the time axis, wave-limited narrow for holds.
• Add reuse windows and the report-header wave fact as new public report surfaces; cache facts remain target facts and appear only in capacity findings.
• Merge ParallelCapacityFacts into TopologyLevelFacts, which states max_logical_units and max_physical_units for each topology level under TopologyFacts.
• Format text-report byte counts through one binary unit ladder while JSON keeps raw integer byte values.

Risk

• Capacity compares one wave with one cache instance; deployments that spread a wave across cache instances are not modeled.
• Residency windows are loop-granular upper bounds on reuse distance, not Falcon-style per-access reuse sources.
• A truncated wave requires a statically stateable mesh order; a scope without one states no space axis, and a window containing such a boundary reports no row.
• Under nested mesh regions a reuse row states its space axis against the innermost mesh; outer execution meshes are not composed into that label.