fix(matmul): budget the transient fifth tile in _gemm_tiled_sync's per-step working set

[RenzoMXD]

PR kind

• fix
• feat

(Replaces this PR's original content -- a nystrom-iter feat idea that a real GPU scorecard showed doesn't clear the dominance gate at the regime tested. Repurposing the branch/PR for unrelated, verified content instead of abandoning it.)

Summary

_tile_workspace_bytes_per_elem budgets four T×T tiles per k-step in matmul/gemm.py's tiled path: acc, a_dev, b_dev, prod. But _gemm_tiled_sync reassigns a_dev, b_dev, and prod every iteration (a_dev = backend.to_device(...), etc.), and Python evaluates the right-hand side of an assignment -- allocating the new tile -- before the name rebinds and the old tile's reference drops. So at the moment any one of those three lines runs, the previous iteration's tile it's about to replace is still live alongside the freshly allocated one: a fifth T×T tile, momentarily, on top of the steady-state four.

Budgeting only four under-counts true peak device usage by one operand-sized tile (roughly 20-25% depending on T), risking OOM at a vram_fraction that leaves little headroom -- exactly the failure mode the tiled path exists to avoid.

I verified the mechanism in isolation with a standalone CPython reference-counting trace before touching this file: steady state was 4 tile-sized objects, measured peak was consistently 5, across fp16/fp32/fp64 and accumulate_fp32 on/off.

Fix

_tile_workspace_bytes_per_elem: cfg.acc_dtype.itemsize + 3 * _tile_operand_bytes(cfg) → ... + 4 * _tile_operand_bytes(cfg). Budget-only change -- the blocked GEMM math (_tiles, the k-loop accumulation) is untouched.

Validation

• uv run python -m strategy.smoke
• uv run --extra test python -m pytest tests/ strategy/tests/ eval/tests/ -q — 288 passed, 24 skipped
• Updated tests/test_auto_tile_budget.py's two tests that hardcoded the old 3-operand-term formula (test_tile_workspace_counts_fp16_upcast, test_tile_workspace_counts_the_bmm_output_tile), and added test_tile_workspace_counts_the_transient_fifth_tile explaining and asserting the new 4-operand-term model directly.

No protected paths touched (matmul/, tests/ are Open per CONTRIBUTING.md).

Checklist

• CPU-safe validation passed
• My commits do not include Co-authored-by footers for coding agents
• Correctness gates pass

[zeokin]

Budgets the transient 5th GEMM tile in _gemm_tiled_sync; fixes a real peak-VRAM undercount.