0.25.9

Added

• The cuda and hip backends now generate faster code for scans
  that have been fused with maps that internally produce arrays.
  Work by Anders Holst and Christian Påbøl Jacobsen.

• f16.ldexp, f32.ldexp, f64.ldexp, corresponding to the
  functions in the C math library.

Fixed

• Incorrect data dependency information for scatter and vjp could
  cause invalid simplification.

• Barrier divergence in certain complicated kernels that contain both
  bounds checks and intragroup scans.