perf(householder): no parallel region below the compact factor's own floor

[shinaoka]

What

upgrade_householder_factor opens a rayon region for its V^H V products whenever the caller asks for threads — and its own deep predicate right beside them already states when that reduction has the shape to pay for it (deeper than 2 * n, at least one row per thread, over the 8 * 1024 work floor). Below that floor the products still opened a region each, and a blocked QR calls this function once per block, at a block size the recursion halves every level. A threaded 128 x 128 QR thus paid the fork-join cost of n-wide reductions about fifty times for work the same floor calls too small to split.

The change asks the floor once, before anything is opened: below it the call takes Par::Seq, above it the caller's par is forwarded exactly as before. Nothing about what is computed changes; the predicate is the one the products themselves already use, so no new constant.

Measured (isolated driver on the call tlinalg's qr family makes: qr_in_place, CPUs 4-11, best of 7)

ms at Par::rayon(8), before → after (f64 unless stated):

shape	before	after	before / seq	after / seq
128x128	1.05	0.235	4.2-4.6x	1.00
192x192	1.79	0.567	2.9-3.0x	0.94
256x256	2.46	0.917	2.1-2.3x	0.79
384x384	3.12	1.85	1.3-1.4x	0.60
512x512	4.59	2.85	0.96-0.98	0.44
768x768	7.32	5.47	0.49	0.38
1024x64	0.65	0.466	1.34-1.44x	1.00
4096x256	6.32	6.31	0.40	0.40
512x512 c64	7.36	5.18	0.42	0.30
256x256 c64	2.53	1.34	0.95	0.50
1024x64 c64	0.99	0.86	0.73-0.76	0.63
4096x256 c64	16.67	16.64	0.30	0.30

Every shape improves or is unchanged. The tall shapes the ledger's other rows serve keep their parallelism (their reductions are above the floor); below the floor no region is opened at all rather than a smaller one.

Two neighbouring candidates were measured and dropped for this PR, both recorded here so they are not re-tried: gating the per-block apply on the trailing area ((m - row) * (n - col) < par_threshold) is harmful on tall panels — their block updates are rows x local x a1cols with a long depth, so an area threshold sends real wins to Par::Seq (4096x256 f64 6.36 → 6.82, c64 16.53 → 18.23) — and forcing the n x n x n head product of the prev_block_size < 8 branch to Par::Seq costs 38% at 512x512 f64 (2.49 → 3.43).

Gate

cargo test -p t4a-faer --lib (202 passed) and --doc (37 passed). One pre-existing failure in the fork is unrelated to this change: sparse::linalg::qr::tests::test_numeric_simplicial_qr_1_transpose panics on an assertion in triangular.rs:990 on the parent commit too (verified on a clean checkout of 22f23cf).

PATCHSET.md carries the row, including the measurement protocol.