fix(docs): make matmul README FLOP table use one (decimal) n convention (Fixes #102)

[luciferlive112116]

Fixes #102.

PR kind

• fix
• feat

Summary

matmul/README.md's "it doesn't fit" table mixes two n conventions within a
single table. The memory columns use decimal n (128k → 128000, so one fp32
matrix is 128000² × 4 = 65 GB, and A+B+C = 196 GB), but the compute (2n³)
column used binary n for the first three rows and decimal n for the
last:

n	documented 2n³	from	should be (decimal n)
1k	2.1 GFLOP	2·1024³ (binary)	2.0 GFLOP (2·1000³)
16k	8.8 TFLOP	2·16384³ (binary)	8.2 TFLOP (2·16000³)
32k	70 TFLOP	2·32768³ (binary)	66 TFLOP (2·32000³)
128k	4.2 PFLOP	2·128000³ (decimal — already correct)	4.2 PFLOP

A binary n overstates 2n³ by (1024/1000)³ = 1.074 (~7%), so the first three
rows were ~7% high and inconsistent with the decimal n the same table's memory
columns use.

Fix

• matmul/README.md — recompute the 1k/16k/32k compute cells from the same
  decimal n the memory columns use: 2.1 → 2.0, 8.8 → 8.2, 70 → 66
  (128k already 4.2 PFLOP). Numbers-only; +3/−3.
• tests/test_matmul_flop_table_units.py (new) — CPU-only arithmetic test that
  derives the compute column from 2·(decimal n)³ (and a witness that the old
  binary-n figures were the inconsistent, larger ones) so the table cannot
  drift back to a mixed convention.

Validation

$ python -m pytest tests/test_matmul_flop_table_units.py -q
  3 passed

$ python -m pytest tests/ strategy/tests/ eval/tests/ -q      # CI-equivalent (no torch → GPU tests skip)
  288 passed, 24 skipped

No protected paths (eval/, docs/, .github/, dashboard/) touched — this
edits matmul/README.md (package doc, open zone).

Result

Documentation-consistency fix; no code path changed, so no accuracy / latency /
VRAM delta.

metric	value
accuracy	unchanged (docs + arithmetic test only)
time complexity	unchanged
latency	N/A
VRAM usage	N/A

[zeokin]

Unifies the matmul README FLOP table to a single decimal-n convention; arithmetic test.