Improved basecases

[cxzhong]

Improved base cases, double→float conversion, and MMHelper simplification

Summary

This PR addresses four items from the TODO list:

1. LUdivine-PLUQ cleanup — unified routine with runtime dispatch
2. FTRTRI base case optimization — blocked base case for better cache locality
3. Double→float conversion — extended from fgemm to ftrsm, ftrmm, ftrsv
4. MMHelper simplification — minimal primary template + AddHelper replacing NeedPreAddReduction

All 43 tests pass with zero warnings. Benchmarks show 3–5× speedup for Modular<int32_t> operations with no regression for Modular<double>.

---

Changes

1. LUdivine-PLUQ (ffpack_pluq.inl, ffpack_ludivine.inl, ffpack.h)

• PLUQ: replaced compile-time #ifdef LEFTLOOKING with a runtime FFPACK_LU_TAG enum
  (LAPACK_CROUT, LAPACK_LEFT_LOOKING, TILE_CROUT, TILE_LEFT_LOOKING), defaulting to LAPACK_CROUT.
• LUdivine: unified the float and double specializations into a single SFINAE-based
  template, eliminating ~120 lines of duplicated code.
• Fixed pre-existing bug: cyclic_shift_row_col used Element_ptr instead of Element in
  an intermediate variable.

2. FTRTRI base case optimization (ffpack_ftrtr.inl)

• Added ftrtri_basecase() with a blocked algorithm (block size 4): splits the matrix into a
  small block and remainder, using ftrmm on blocks instead of single rows.
• The row-by-row micro base case is retained for N ≤ 4.
• The recursive ftrtri() now delegates to ftrtri_basecase() at the threshold.

3. Double→float conversion (fflas_ftrsm.inl, fflas_ftrmm.inl, fflas_ftrsv.inl)

• Added _try_convert SFINAE helpers in FFLAS::Protected to attempt conversion from
  Modular<double> to Modular<float> when p² < 2²⁴ (single-precision safe).
• Applied to ftrsm, ftrmm, and ftrsv — matching the existing pattern in fgemm.

4. MMHelper simplification (fflas_helpers.inl, fflas_fgemm.inl, + 3 schedule files)

• Minimal primary template: MMHelper now contains only recLevel and parseq by
  default, suitable for non-bounded modes (DefaultTag, ConvertTo, etc.). The previous
  DefaultTag and ConvertTo partial specializations are removed (now redundant).
• MMHelperBounded base class: all bounds tracking machinery (Amin/Amax, Bmin/Bmax,
  Cmin/Cmax, Outmin/Outmax, FieldMin/FieldMax, MaxStorableValue, delayedField,
  MaxDelayedDim, setOutBounds, checkA/B/Out, etc.) extracted into this base.
• Bounded specializations: MMHelper<..., LazyTag, ...>, MMHelper<..., DelayedTag, ...>,
  and MMHelper<..., DefaultBoundedTag, ...> inherit from MMHelperBounded via using Base::Base.
• IsBoundedMode<ModeTrait> trait: identifies bounded mode categories at compile time.
• AddHelper<IsSub>: replaces NeedPreAddReduction (IsSub=false) and
  NeedPreSubReduction (IsSub=true) free functions. 29 call sites updated across
  schedule_winograd.inl, schedule_winograd_acc.inl, and fflas_fsyrk_strassen.inl.
  NeedDoublePreAddReduction, NeedPreScalReduction, and NeedPreAxpyReduction are kept
  as-is (different signatures).

Bug fixes

• ffpack_permutation.inl: fixed cyclic_shift_row_col using Element_ptr where Element
  was needed.

Files changed (14 files, +620 −374)

File	Description
ffpack/ffpack.h	Added FFPACK_LU_TAG enum
ffpack/ffpack_pluq.inl	Runtime LU variant dispatch
ffpack/ffpack_ludivine.inl	Unified float/double via SFINAE
ffpack/ffpack_ftrtr.inl	Blocked ftrtri_basecase + inline comments
ffpack/ffpack_permutation.inl	Bug fix in cyclic_shift_row_col
fflas/fflas_ftrsm.inl	Double→float conversion
fflas/fflas_ftrmm.inl	Double→float conversion
fflas/fflas_ftrsv.inl	Double→float conversion
fflas/fflas_helpers.inl	Minimal primary MMHelper + MMHelperBounded base
fflas/fflas_fgemm.inl	AddHelper<IsSub> replacing NeedPre*Reduction
fflas/fflas_fgemm/schedule_winograd.inl	Updated 14 call sites
fflas/fflas_fgemm/schedule_winograd_acc.inl	Updated 4 call sites
fflas/fflas_fsyrk_strassen.inl	Updated 11 call sites
TODO	Marked items as done

---

Testing

• make check -j 16: 43/43 PASS, 0 FAIL, 0 warnings
• Benchmarked ftrtri, fgemm, ftrsm on Modular<int32_t> and Modular<double>:
  3–5× speedup for integer modular types, no regression for floating-point types.