perf: per-subgraph k optimization for split schedules

[melroyanthony]

P2: Per-subgraph k optimization

Problem

With cost-based fusion (P1), the schedule may have multiple subgraphs. Each subgraph should independently optimize its k dimension based on its own MatMul ops' K_full values and memory constraints.

Proposed Approach

For each subgraph after fusion decisions:

1. If subgraph contains MatMul ops, search k from min(K_full) down to 1
2. For each k candidate, compute working set and latency
3. Select the k that minimizes latency while fitting in memory (accounting for retained tensors)

Expected Impact

2-10x improvement combined with P1, especially on benchmarks with large K_full values (9, 13: K_full=4096).

Acceptance Criteria

• Each subgraph independently searches for optimal k
• k candidates: K_full, K_full/2, K_full/4, ..., 1
• Working set includes retained tensors from previous subgraphs
• Track A (Rust) and Track B (Python) both updated
• All tests pass

Dependencies

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

[melroyanthony]

Post-P0 Update (PR #20 merged)

This issue is largely resolved by PR #20:

• granularity.rs already searches k from K_full down to 1 in powers of 2
• Each subgraph independently optimizes its (w, h, k)
• The K_full consistency invariant ensures k is well-defined per subgraph

Remaining work (if any)

• Verify k > 1 is actually selected for MatMul subgraphs on the released benchmarks
• The k search may be redundant with the split-K optimizer stage — consider whether both are needed

This issue may be closeable once #16 (performance fix) is done and benchmarks can be verified.