feat: implement Reduced Matrix Multiplication (arXiv:2608.13426)

[erkinalp]

Input-adaptive contraction-axis reduction for the products that dominate inference: Y = A B keeps only the ceil(rho*d) slices with the largest activation column norms ||A[:,j]||_2, chosen deterministically from the current activations. Training-free and weight-preserving.

• ggml-rmm.h: column-norm scoring, quickselect TopK with lowest-index tie breaking, reduced GEMV/GEMM, projections, attention with QK^T feature reduction and optional PV token reduction, reusable workspace, MAC stats. Scalar C11 everywhere, VSX/AltiVec fast path on POWER8, dcbt prefetch of only the retained weight rows. rho = 1 is bit-identical to dense.
• tests/rmm_test.c: TopK determinism and ties, dense equivalence at rho = 1, Proposition 1 error bound, activation-aware vs random selection, minimax claim, attention accuracy at rho_d = 0.5, MAC accounting, GEMV benchmark.
• Makefile: 'make' runs the suite on x86-64, 'make power8' builds the VSX path.
• docs/RMM.md, README and CHANGELOG entries.

[Scottcjn]

Received, and queued for review — apologies for the silence up to now.

This is a maintainer-side backlog rather than any verdict on your contribution. We
had a large number of open PRs with no reply on them at all, which is our failure to
manage, not a signal about your work. Yours is in the queue and will get a real
review with a real decision.

I'm not going to give you a date I might miss. What I can tell you:

• Reviews run oldest-first.
• If this PR has gone stale against main, say so and I'll prioritise closing it so
  the queue reflects what you actually still want landed.
• If you're submitting this for bounty credit, link the bounty issue in a comment —
  several contributions are eligible and we'd rather pay you than have you wonder.

Thanks for your patience, and for the work.