fix(strategy): nystrom uses 3 landmark spaces, not 4 (drop redundant col(B))

[galuis116]

Summary

Fixes #269.

NystromTransform split its M-column landmark budget across four spaces —
col(A), row(A), col(B), row(B) — but col(B) is redundant. With the
projector P = Q Qᵀ, Ĉ = P A P B P = A B once range(Q) contains only
col(A), row(A), row(B):

P A = A  (col(A) ⊆ range(Q)),  A P = A  (row(A)),  B P = B  (row(B))
=>  P A P B P = A B P = A B

col(B) contributes nothing — the exact argument rsvd already uses (3 sketches,
#91 "rsvd wastes ~25% of the rank budget"). nystrom never got it, so a full
M/4 of the budget was wasted, and exact recovery of a rank-r product needed
M ≳ 4r instead of 3r — ~25% worse accuracy per M on the low-rank /
decaying-spectrum regimes it targets.

The fix splits M across the three necessary spaces (drops the col(B) block).
Same basis cost (gather M columns + thin QR — latency/VRAM unchanged),
strictly higher accuracy per M.

Test plan

• python strategy/tests/test_nystrom_three_spaces.py — 3/3 (CPU): recovers a
  rank-4 A@B at M=15 (≈5 landmarks/space); basis orthonormal & correctly shaped.
  This test fails on the current 4-space nystrom (M/4 = 3 < r = 4 per space →
  cannot recover) and passes on the fix — a direct demonstration of the wasted budget.
• ruff check strategy/transforms.py strategy/tests/test_nystrom_three_spaces.py — clean.

Notes

• Distinct from feat(strategy): rrqr-nystrom -- redundancy-avoiding landmark selection for decaying-spectrum data #226, which adds a new rrqr-nystrom transform and does not
  touch the existing NystromTransform; this fixes the existing one.
• Also brings NystromTransform.basis's signature in line with the base class /
  rsvd (frac=None), completing the frac-signature consistency of [bug] transform_template's basis() still omits frac, so every copied transform silently ignores --vram-fraction (the #211 bug, re-introduced per contribution) #250–fix(strategy): finish the frac basis() signature across contributor-facing snippets (Fixes #252) #253
  (nystrom gathers on the host and doesn't stream, so frac is simply accepted).

No scorecard: this is a correctness/efficiency fix: to a built-in transform (the
budget split), not a new strategy — fix: PRs are scorecard-exempt in the gate chain.
The change only reallocates existing landmark budget, so latency/VRAM are unchanged
while accuracy-per-M strictly improves.