Shorten backward-substitution dependency chains

[ChrisRackauckas-Claude]

⚠️ Draft — please ignore until reviewed by @ChrisRackauckas.

Follow-up to #41. Performance-only; no API or semantics change; forward-substitution kernels untouched.

What changed and why

Per-solve hardware counters (isolated by differencing two iteration counts) showed the backward family executing nearly the same instructions as its forward twins but at less than half the IPC — e.g. at n=12 (a single register-blocked call, zero reduction trips): forward 1785 insns / 543 cycles (IPC 3.3) vs backward 1825 insns / 1208 cycles (IPC 1.5). Pure dependency stalls.

Root cause: solve_AU (and the forward small kernels) emit each column's updates with ascending k, so the freshest operand (A_{n-1}) arrives last — older updates overlap its latency and the critical path is one fma per column. The backward kernels emitted for k = n+1:N ascending, which puts the freshest operand (A_{n+1}) first, serializing the remaining N-n-1 fmas behind it: an O(W²/2) chain per diagonal-block solve instead of O(W). The same inversion existed in BdivL_small_kern[_u]! and in uldiv_solve_W_u!'s cross-sub-block j-loop.

Fix: emit those reductions with descending k/j (freshest last). Also re-based the backward reduction loops to zero-based induction with pre-offset (gesp) pointers, matching the forward kernels' loop shape.

Measured (EPYC 7502 / AVX2, Float64, nrhs=8, min times, 1 BLAS thread)

Backward vs forward at equal UNIT flag (ldiv! transposed family, μs):

   n  | uLower  uUpper(before) uUpper(after) | ratio before -> after
   48 |   0.68     0.96           0.78       |  1.39x -> 1.15x
   64 |   1.24     1.62           1.38       |  1.30x -> 1.11x
  128 |   4.45     5.20           4.70       |  1.17x -> 1.06x
  256 |  19.37    20.95          20.31       |  1.09x -> 1.05x
  500 |  73.77    77.40          75.43       |  1.05x -> 1.02x
 1000 | 281.57   290.25         286.08       |  1.02x -> 1.02x

Left-upper ldiv! (non-unit) vs OpenBLAS trsm, same protocol:

   n  |   BLAS   TS upper   speedup
   32 |   1.63     0.64      2.55x
   64 |   4.09     1.54      2.66x
  128 |  12.36     5.07      2.44x
  256 |  48.20    20.24      2.38x
  500 | 151.81    76.93      1.97x
 1000 | 581.25   291.33      2.00x

(#41's table had 1.9–2.3x on these points.)

Verification

• Full test suite: TriangularSolve.jl | 148781 148781 pass (Julia 1.12.4) plus Aqua/ambiguities; extended sweep (n=1..400 all remainder cases, nrhs=1..32, Float32/64, unit/non-unit, 2/3-arg, threaded, packed lu! parents to 512, BigFloat fallback): 0 failures; zero steady-state allocations.
• Forward kernels: code untouched by this diff; spot re-bench consistent with Add native left-upper ldiv! and right-lower rdiv! kernels #41 within noise.
• Not verified: AVX-512/aarch64 (AVX2 machine), MKL.

Note on versioning: 0.2.2 (from #41) is not yet registered — registration needs a JuliaSIMD org member to comment @JuliaRegistrator register (my bot account was refused). If this merges first it can ship inside 0.2.2; if 0.2.2 is registered first, this should bump to 0.2.3.

🤖 Generated with Claude Code

[JuliaRegistrator]

Error while trying to register: Register Failed
@ChrisRackauckas-Claude, it looks like you are not a publicly listed member/owner in the parent organization (JuliaSIMD).
If you are a member/owner, you will need to change your membership to public. See GitHub Help

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 94.55%. Comparing base (258a420) to head (b28192b).
⚠️ Report is 13 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #44      +/-   ##
==========================================
+ Coverage   85.01%   94.55%   +9.54%     
==========================================
  Files           1        1              
  Lines         427      827     +400     
==========================================
+ Hits          363      782     +419     
+ Misses         64       45      -19     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.