perf: tensor retention optimization between subgraphs

[melroyanthony]

P3: Tensor retention between subgraphs

Problem

With multiple subgraphs (from P1), boundary tensors are written to DRAM then re-read. Retaining tensors in fast memory across subgraph boundaries eliminates this round-trip cost (`2 * tensor_size / bandwidth`).

Current State

Retention logic exists (`optimizer/retention.rs`) but is never triggered because the mega-fusion strategy produces only 1 subgraph with 0 boundaries.

Acceptance Criteria

• After fusion decisions create multiple subgraphs, retention pass identifies candidates
• Retention only applied when residual capacity allows (retained tensor at full size)
• Net latency improvement validated: retained saves > capacity cost
• Track A (Rust) and Track B (Python) both updated
• Verified against Example 3C pattern (4,638.4 latency)

Dependencies

Depends on #16 (cost-based fusion) — only relevant with multiple subgraphs.

[melroyanthony]

Post-P0 Update (PR #20 merged)

The K_full consistency check now creates multiple subgraphs, which means retention is now active in the pipeline. optimizer/retention.rs runs as stages 3 and 6.

Remaining work

• Verify retention is actually improving latency on the released benchmarks (need perf: cost-based fusion instead of greedy merge-everything #16 perf fix first to run benchmarks to completion)
• The retention logic may need tuning for the new subgraph structure
• Consider retention in the cost-based fusion decision (perf: cost-based fusion instead of greedy merge-everything #16): the boundary cost of splitting should account for potential retention savings

May be closeable once benchmarks are verified with retention working correctly.