Fix minimum-norm least squares for rank-deficient inputs - #12613

Fixes #12610.
The QR-based least-squares implementation unconditionally uses a triangular solve. Rank-deficient coefficient matrices can raise a singular-matrix error or produce unstable coefficients instead of a minimum-norm solution. Computing singular values from RᴴR also loses numerical accuracy by squaring the spectrum.
Changes:

• Retain distributed QR and compute one shared reduced SVD of the small R block
• Use the original matrix dimensions and QR precision for the solution/rank cutoff
• Preserve residual norms, RHS column chunks, complex conjugation, and scalar int64 rank metadata
• Compute unknown row counts through scalar shape tasks without gathering the full input matrix
• Handle real and complex subnormal scales without unstable singular-value reciprocals
• Make unsupported coefficient layouts explicit
  Validation:
• All 452 linalg tests passed on both current and minimum dependency environments
• Independent numerical and graph review passed
• Configured pre-commit hooks passed for changed files
• Array suite: 4,915 passed, 551 skipped, 16 xfailed; two memory-heavy tests excluded
• Whole-project collection was blocked by missing optional pyarrow; no full-project pass is claimed
• GPU/CuPy execution was not tested