Fix RowMajor performance for triangular/dense assignment

Summary

• Fixes #3031 (closed)
• Rewrite the dynamic triangular_assignment_loop to iterate in storage order (outer/inner matching layout) instead of always iterating outer=col, inner=row
• This gives contiguous memory access for both ColMajor and RowMajor storage, fixing a 5-137x RowMajor performance deficit while maintaining ColMajor parity
• Use compile-time constexpr row()/col() helpers that constant-fold to zero overhead
• Keep simple scalar loops so GCC recognizes memcpy/memset idioms and Clang auto-vectorizes

Benchmark Results (float, GCC/Clang x Haswell/Westmere)

RowMajor Triangular2Dense (GCC Haswell):

Size	OLD	NEW	Speedup
64	1091 ns	199 ns	5.5x
256	206,778 ns	2,423 ns	85x
1024	7,830,936 ns	57,028 ns	137x

ColMajor - near-parity across all configs (0.93x-1.36x, median ~1.00x).

Test plan

• All triangular_1..5 tests pass
• All selfadjoint_1..5 tests pass
• All product_trsolve_1..5 tests pass
• Benchmarked across GCC/Clang x Haswell/Westmere (4 configs, 3 repetitions)
• Verified assembly: GCC uses memcpy/memset, Clang auto-vectorizes with vmovups
• clang-format clean

Generated with Claude Code